Shared Hermes chat links now give people more than a single message thread. The chat page has a sidebar for switching conversations, seeing what the assistant can do, and reaching owner settings when you have access.

• Chat sessions in the sidebar: Start a fresh chat, return to recent conversations, and keep each shared link organized on the device where it is being used.
• Skills are visible from chat: The Skills view shows the assistant's installed capabilities with descriptions, so users can understand what the assistant is equipped to handle before they ask.
• Owner controls stay close: Assistant owners can jump from the shared chat to Secrets, while public visitors only see the controls that make sense for them.
• Cleaner shared chat UI: The chat header moved into the sidebar, theme controls are easier to find, and the page keeps more room focused on the conversation.
• More dependable responses: Slow or interrupted Hermes replies now fail more clearly instead of leaving the chat stuck waiting.

This makes shared Hermes links easier to hand to customers, teammates, and collaborators. They can chat, understand the assistant's skills, and keep separate sessions without needing the full owner dashboard.

WhatsApp assistants now have more precise reply rules. You can keep one default instruction for everyone, then add overrides for the exact people or groups where Hermes should behave differently.

• Everyone, numbers, and groups: The Rules section now separates the global default from specific contacts and WhatsApp groups, so one assistant can handle different conversations with different guidance.
• Contact-specific instructions: Add a number or pick a WhatsApp contact, then tell Hermes how to answer that person. That means VIP clients, suppliers, or teammates can get tailored behavior without changing the default rule.
• Group-specific instructions: Give busy groups their own policy, like answering only direct requests for help and ignoring casual chatter.
• WhatsApp search in setup: Search contacts and groups from the Rules panel instead of copying chat ids by hand. Recent inbox chats still help when a live address book is warming up.

This makes WhatsApp automation easier to use in real operations. Hermes can stay broad by default, then get stricter, quieter, or more specific exactly where the conversation needs it.

WhatsApp auto-reply is getting safer for real customer conversations. You can now define what Hermes must not disclose, tune how much of its work is visible in chat, and manage pairing and automation from a clearer WhatsApp setup flow.

• Disclosure guardrails: Add rules for private details Hermes should never reveal, like internal supplier names, payment details, customer phone numbers, private calendar notes, or unpublished discounts.
• Reviewed replies before delivery: When guardrails are active, drafted WhatsApp replies are checked before they are sent. If a draft crosses the line, Hermes rewrites it or falls back to a safe refusal.
• Cleaner customer-facing messages: Turn off the visible [hermes] marker when you want auto-replies to read like normal WhatsApp messages.
• Optional thinking steps: Choose whether contacts see Hermes tool-progress bubbles while it works, or only the final answer.
• Clearer WhatsApp setup: Pairing, message output, reply scope, instructions, allowlists, and guardrails now live in a more structured WhatsApp panel, so it is easier to see what is ready and what still needs attention.

This makes WhatsApp assistants easier to use in higher-trust conversations. You can keep Hermes helpful while controlling what it says, how it appears, and when it stays quiet.

Some hosted apps are meant for a specific client or team. Others are meant to be opened by anyone with the link. You can now mark a Hermes-hosted app as public directly from the Apps tab, while keeping the rest of your assistant's apps private or limited to the exact people you choose.

• Public hosted apps: Use the new public option when a dashboard, report, website, or demo should open without sign-in. That means customer-facing work can be shared as a normal link.
• Private apps stay protected: Owner-only, email-specific, and company-domain access still work for the apps that should stay internal.
• Clearer access state: App cards now show when an app is public, and reopening the access panel keeps that state visible instead of making public apps look owner-only.
• Safer public switching: Public access is synced with the edge protection in front of the app and with the assistant's own app registry, so the link behavior and dashboard state stay aligned.
• Healthier Hermes gateways: Active Hermes assistants are checked in the background, and unhealthy gateways are repaired automatically when the assistant is not already deploying or updating.

This makes hosted Hermes apps easier to use as real deliverables. You can publish something private, share it with a narrow audience, or make it fully public without changing how the assistant itself is managed.

You can now decide who gets into each hosted app your Hermes assistant publishes. A private dashboard can stay owner-only, while a client report or team page can be opened to the exact emails or domains you choose.

• Per-app sharing controls: Every hosted app gets its own Access panel in the Apps tab, so one dashboard can be shared with a client while another stays private.
• Invite by email: Add individual people when you want a small, exact audience for a dashboard, report, or internal tool.
• Allow a whole domain: Share with everyone at a company by adding a domain like example.com, without typing every teammate by hand.
• Owner access stays protected: The assistant owner always keeps access, even when the shared audience changes.
• Clear visibility in the Apps tab: App cards now show whether an app is owner-only or shared, so it's easier to audit what is exposed.

This turns hosted Hermes apps into something you can confidently share. You still get protected links, but now the audience can match the work instead of being all-or-nothing.

Hermes assistants can now turn generated work into live dashboards and websites. Instead of stopping at files in a session, your assistant can publish a protected app that stays attached to the assistant and shows up in GetClaw.

• Publish from the assistant: Ask Hermes to create or update a dashboard, report, landing page, or small website and it can make that work available through the assistant's own live link.
• New Apps tab: Every Hermes assistant now has an Apps tab that lists published surfaces with names, descriptions, update times, and a direct Open button.
• Protected by default: Published apps use the same sign-in protection as the assistant, so private dashboards and internal pages do not become public just because they are easy to open.
• No extra hosting setup: Hermes can publish the finished work itself. You do not need to connect Vercel, Netlify, Cloudflare, or a separate server just to view the result.
• Built for living dashboards: Published apps can be refreshed as the assistant gathers new information, so a report can become a reusable dashboard instead of a one-off artifact.

This makes Hermes more useful after it finishes the first task. The output can become a place you return to, share with the right people, and keep improving with the same assistant.

WhatsApp auto-reply is now more useful in real conversations. Instead of relying on a fixed trigger word or answering every eligible chat the same way, you can write plain-language instructions that tell Hermes when it should answer and when it should stay quiet.

• Custom auto-reply instructions: Add instructions like “answer only when someone is trying to book a dining slot” directly from the WhatsApp panel. Hermes uses that rule before sending anything back.
• Quiet by design: If a message does not match your instructions, Hermes sends no WhatsApp reply. That means you can connect busier chats without creating noise.
• All DMs + groups is smarter: The group-chat mode now evaluates inbound DMs and group messages with your instructions instead of requiring people to start every message with a special prefix.
• Default modes still work: Self and allowlist modes keep their familiar behavior. Leave the instructions blank when you want the normal generic auto-reply, or add guidance when you want more control.
• Visible typing feedback: WhatsApp now shows that your assistant is typing while Hermes is working, then clears the indicator when the reply is done or when Hermes decides not to answer.

This makes WhatsApp feel less like a bot switch and more like a real assistant policy. You decide the situations where Hermes should participate, and the assistant stays out of everything else.

Until today, your WhatsApp assistant only worked in direct messages, and the “All DMs” mode answered every single incoming chat. Useful for solo testing, noisy in real life. Group chats were off-limits entirely. Both of those things change with a small new convention: the word gg.

• Group chats are in: Add your assistant's WhatsApp account to any group and it'll reply on demand. It stays quiet unless someone explicitly calls it.
• gg is the trigger: Type gg followed by your question, like gg what's on my calendar today?, and only that message reaches Hermes. Anything without the prefix is ignored, so normal group chatter doesn't spawn replies.
• Call it from your own phone too: When you're sitting in a group on the phone your assistant is paired to, type gg ... yourself and the reply lands right in the group. Great for asking your assistant in front of the people you're planning with.
• Case-insensitive, prefix stripped: gg, GG, and Gg all work. The prefix is removed before Hermes sees the message, so your prompts stay clean and the answers don't echo the trigger word back.
• Spelled out in the WhatsApp panel: The auto-reply mode is now called All DMs + groups. Pick it on any WhatsApp-connected assistant and an inline explainer shows the trigger word, a working example, and exactly which messages will be ignored.
• Self and Allowlist modes are unchanged: If you'd rather keep your assistant scoped to your own “Me” chat or a hand-picked list of contacts, those modes still work the way they did, with no prefix needed and groups still off.

Groups have been on the wishlist since the day we shipped WhatsApp. Now you can drop your assistant into a planning thread, a family chat, or a small team group, and pull it into the conversation only when you actually want it to chime in.

Until now, talking to your Hermes assistant meant going through a chat channel like WhatsApp, Telegram, or iMessage. There was no clean way to put it inside your own product. A new public Chat API changes that. Every Hermes assistant gets its own OpenAI-compatible endpoint, scoped to that assistant, that you can call from any app you build.

• Drop-in for any OpenAI SDK: The endpoint speaks the same shape as OpenAI's chat completions API. Set your getclaw API key as the SDK's key, point the base URL at your assistant, and the rest of your code stays exactly the way you wrote it.
• Stream replies token by token: Responses come back over a streaming connection, so your UI can render the reply as it's being generated. Cancel from your side and the assistant stops working on it too, no wasted tokens.
• Authenticated with your getclaw API key: The same key you already use to manage assistants now unlocks the Chat API. Send it as a Bearer token or an X-API-Key header. No separate keys to juggle.
• Server-side memory per end user: Pass an X-Hermes-Session-Id header and your assistant keeps the conversation going across calls, on its side, scoped to that id. Skip the header for stateless one-shot calls. Either pattern works without you running a database.
• A new Chat API tab on every Hermes assistant: Open any Hermes assistant in the dashboard and you'll find a Chat API tab with copy-paste curl, OpenAI SDK, and session-aware snippets pre-filled with your deployment. Grab a working request in seconds.
• Documented in the API reference: New entries for chat completions and the model list show up in the API reference and the OpenAPI spec, so the same docs your team already reads now describe the chat surface too.

The whole point of giving your assistant skills, memory, and channels is to make it useful in the places your users already are. Now one of those places can be your own product, behind your own UI, with the assistant doing the work.

If you'd rather not register a number with a third-party iMessage provider, you can now point Hermes at a BlueBubbles server running on a Mac you own. Replies go out from your real iMessage account, on the device you already use, with nothing else in the middle.

• Use your own iMessage account: Replies come from the Apple ID signed into your Mac. No new number, no shared shortcode, no extra middleman holding your conversations.
• One panel, one webhook: Open the iMessage via BlueBubbles tile on any Hermes assistant, paste your server URL and password, and copy the webhook URL we generate straight into BlueBubbles. The setup guide walks you through every step.
• You decide who it answers: Start with an allowlist of specific chats and grow it over time, or open it to every contact at once. Anyone outside the list gets nothing back.
• Pause without unpairing: A single switch on the panel turns auto-reply off while keeping your connection live. Flip it back on whenever you're ready and your assistant picks right up.
• Server password stays encrypted: We store your BlueBubbles password encrypted and never echo it back to the page. Rotate it any time by entering a new one, and the same webhook URL keeps working.
• Disconnect anytime: One click clears the credentials, stops your assistant from answering, and leaves your Mac and your threads exactly where they were.

For people who already keep a Mac running at home or in the office, this is the most direct way to put an assistant on iMessage: your account, your hardware, your control, with the convenience of a managed setup flow on top.

Until now, the only env vars your assistant could see were the ones we set up ourselves. Bringing in a Stripe key, a Notion token, or any other API credential meant editing files inside the container by hand and crossing your fingers it survived the next redeploy. A new Secrets tab on every Hermes assistant fixes that.

• See what's already injected: The Platform-managed list shows every variable we set on your behalf, with a short note on what each one does. Empty slots and internal plumbing are hidden, so you only see things that actually matter for your assistant.
• Reveal any value on click: Values are masked by default. Hit Show on any row to see the real key, token, or URL, and it auto-hides again after a few seconds so it doesn't sit on your screen.
• Add your own keys for skills: Drop a custom variable into the Your Variables list and your skills can read it as a normal environment variable from inside the container. Names like STRIPE_API_KEY or NOTION_TOKEN are yours to define.
• Stack changes, redeploy once: Click Add Variable as many times as you need; each click drops a new draft row in the table. Fill them all in, then hit Redeploy once and they all go live together. No more one-redeploy-per-key.
• Edit and remove without leaving the page: Update a value or pull a custom variable straight from the table. Changes batch up alongside your drafts so you only push to the live assistant when you're ready.
• Encrypted at rest, persistent across redeploys: Anything you add is encrypted in your account and re-injected on every boot. Custom variables stick around through redeploys, region moves, and model changes without you having to re-enter them.
• Safe by default: Names that collide with the variables we manage for you are rejected up front, so you can't accidentally clobber your provider key or a channel token by typing the wrong thing.

Most of what makes an assistant useful in real life is the services it can reach. Now you can hand yours any credential it needs, see exactly what it's working with, and change your mind whenever, all from one tab.

iMessage is now a channel for Hermes assistants, powered by Sendblue. Register a number with Sendblue, paste the credentials into the dashboard, point Sendblue at the webhook we generate, and inbound iMessages start flowing to your assistant. Replies go back through the same number, with the blue bubble.

• Set it up in one panel: Open the iMessage tab on any Hermes assistant, drop in your Sendblue API key, secret, and registered number, and copy the webhook URL and secret straight into Sendblue. The whole loop fits on one screen with a step-by-step guide right there.
• Bring your own number: The number that texts your contacts is the one you registered with Sendblue, not a shared shortcode or a stranger's line. People see a normal blue bubble from a number you control.
• You decide who it answers: Start with an allowlist and add the specific chats your assistant should reply to. Open it up to everyone in your contacts later, or keep it tight forever. Anyone outside the list gets nothing.
• Toggle auto-reply without unpairing: One switch on the iMessage panel pauses replies while keeping the connection live. Flip it back when you're ready and your assistant picks right up.
• Credentials stay encrypted: Your Sendblue API secret is stored encrypted and never echoed back to the page. Rotate it whenever you want by entering a new one and saving.
• Disconnect anytime: One click clears the credentials and stops the assistant from answering. Sendblue keeps your number, your contacts keep their threads, and nothing else changes.

iMessage is where a lot of personal and small-team conversations actually happen. Putting your assistant in that thread, behind a number you own and a list you control, makes it reachable in the place people already text from.

The interactive shell on every Hermes assistant page has a new emulator under the hood. It's the same engine behind Ghostty, the modern terminal, ported to run directly in your browser. Same Connect, Reconnect, and Clear buttons you already use, with a much better experience once you're typing.

• Powered by Ghostty: The emulator now handles escape sequences, Unicode, and color codes the way a serious native terminal does. Output that used to come back garbled or half-cleared now lands cleanly.
• TUI apps just work: Run vim, htop, nano, less, or any ncurses-based tool inside your container and it renders correctly, redraws cleanly, and gives back a tidy prompt when you exit.
• Real text selection and copy: Drag to select, double-click to grab a word, Cmd+C to copy. The terminal renders to the page like ordinary text, so the selection behavior you already know works here too.
• Smoother resizing: Resize your browser window and the terminal reflows immediately, telling the running shell its new dimensions so prompts and TUIs stay aligned without a manual redraw.
• Same controls, no migration: Connect, Reconnect, Clear, and Disconnect all live exactly where they did. The change is entirely under the hood — open a session and you'll feel the difference, not see a new UI.

The container terminal is the place you reach for when something isn't behaving and you want to look at it directly. Making it accurate, native to the page, and out of your way matters more than it gets credit for.

WhatsApp is now a first-class channel for Hermes assistants. Pair it once from the dashboard and the assistant lives inside your existing WhatsApp account, replying to the people and threads you trust without anyone needing to install anything new.

• Pair with a QR code: Open the WhatsApp tab on any Hermes assistant and scan the QR with your phone, the exact flow you already use for WhatsApp Web. The session stays linked until you unpair it, so you only do this once.
• No Business account, no extra number: Your assistant uses your real WhatsApp account. There's nothing to register with Meta, no separate SIM to buy, and no approval queue to sit in before you can talk to anyone.
• You decide who it answers: An allowlist on the assistant page controls exactly which contacts and groups the assistant replies to. Everyone else sees nothing different, so your normal chats stay normal.
• Handles photos, voice notes, and documents: Send your assistant an image, a voice memo, or a PDF and it can read, transcribe, or work on it just like a text message. It can also send media back, so replies aren't stuck in plain text.
• Talks to itself when you want it to: Message your own number and the assistant treats it as a private command line. Great for quick captures, reminders, and one-off tasks without spinning up a separate chat.
• Unpair anytime: One click on the dashboard ends the session and removes the assistant from your WhatsApp account. Your messages and contacts stay where they always were.

WhatsApp is where most of the world already messages. Putting your assistant there means people can reach it without learning a new app, downloading anything, or changing how they talk to you.

Until now, every assistant ran in whichever region happened to be closest to wherever it first spun up. You can now choose the region yourself from the assistant's configuration, with nine options spanning the major continents.

• Pick from nine regions: Eastern and Western North America, Western and Eastern Europe, Asia Pacific, South America, the Middle East, Oceania, and Africa. The picker calls out regions with limited capacity so you know what to expect.
• Defaults to Eastern North America: New assistants come up in Eastern North America by default. If that's where most of your users or data already live, you don't need to do anything.
• Change it whenever you want: Open your assistant's details page, hit the pencil next to Region, pick a new one, and save. If your assistant is live, a redeploy starts automatically so the new region takes effect.
• Better latency for non-US users: If your assistant chats mostly with people in Europe or Asia, moving it closer to them shortens the round trip on every message and tool call.
• Keeps traffic close to your data: Calling APIs or databases hosted in a particular region? Run your assistant in the same one and you save a hop on every request.

Latency adds up over a long conversation, especially when tools are involved. Putting the assistant in the right place is one of the easiest ways to make it feel snappier without changing anything else.

Every assistant inbox now has a handle you control. Instead of the auto-generated address you got when email was first enabled, you can pick something short and memorable and hand it out with confidence. The change applies to both OpenClaw and Hermes assistants.

• Choose your own handle: On the Email tab for any assistant, type the name you want before the @ and save. Your inbox address updates immediately and any future email sent by the assistant goes out from the new address.
• Change it whenever you want: Renamed your assistant, or just want something catchier? Update the handle again and the new address takes effect right away. No redeploy to babysit on OpenClaw, and a seamless redeploy on Hermes.
• Globally unique, guaranteed: Handles are reserved the instant you claim one, so two assistants can never share an inbox. If someone else has already taken the name you want, you'll see a clear message and can try another.
• Friendly validation as you type: The form tells you right away if a handle is too short, uses unsupported characters, or touches one of the reserved mailbox names. No guessing, no trial-and-error saves.
• Defaults unchanged if you don't care: Assistants that already have email enabled keep their existing address — nothing moves until you choose a new handle.
• A cleaner delete confirmation: Deleting an assistant now opens a proper in-app dialog instead of a browser prompt, with a loading state while the tear-down runs and a smooth exit animation on the assistants list.

A random identifier is fine for an internal ID, but it's a terrible thing to hand out to colleagues, customers, or vendors. Now your assistant's inbox can look like it belongs to a person, not a database row.

Hermes is now more than 50% faster. Below is time to first reply — how long between a user's message and the assistant's first word back — tracked over the last few weeks. The baseline held steady for a long time; this release is the cliff on the right.

Two shifts landed in this release. Replies are dramatically faster (above), and assistants now keep their full state across redeploys — every conversation, uploaded file, scheduled job, and approved pairing survives.

• Over 50% faster replies: Simple questions that used to take about 30 seconds now come back in around 7. Complex tasks that took a full minute now finish in roughly 20.
• Automatic snapshots on redeploy: The moment you trigger a redeploy, your assistant's state is captured and saved before the running instance is replaced. Nothing extra to configure, nothing to remember.
• Seamless restore on the new instance: When the replacement instance comes up, it boots with the exact workspace, memory, and channel pairings the previous one had. Conversations pick up right where they left off.
• Existing assistants migrated automatically: Hermes assistants created before this change have their prior data moved over into the new snapshot format, so long-running assistants keep their history without any manual work.
• More forgiving redeploys: A redeploy no longer fails just because your assistant is busy working through a backlog or waiting on an external provider. The platform gives it room to settle instead of restart-looping.
• Faster channel reconnection: After a redeploy, Telegram pairings and other approved channels are already restored, so there's no re-pairing or re-authorization step.

Redeploys used to be the scariest button in the product for long-running Hermes assistants, and waiting on replies was the slowest part of a typical day. With this release, shipping an update is routine instead of a data-loss risk, and your assistant answers while you're still looking at the screen.

When you create a managed assistant, the model picker now pulls in the full OpenRouter catalog instead of a short curated list. That means you can browse and search across 500+ models from dozens of providers, all in the same dropdown. We also updated the defaults so new assistants start on the latest models.

• Full model catalog: The model selector during onboarding now shows every model available through OpenRouter, organized by provider. Popular providers like Anthropic, OpenAI, Google, and DeepSeek appear first, and you can type to search the rest.
• Instant search across all models: Start typing a model name, provider, or ID and the list filters down immediately. No more scrolling through a fixed list hoping the model you want is there.
• GPT-5.4 Mini as the new default: Managed assistants now default to OpenAI's GPT-5.4 Mini, which offers strong performance at a lower price point. You can always change it before confirming.
• GPT-5.4 and GPT-5.4 Mini added: Both models are now available for managed and bring-your-own-key assistants, so you can use them regardless of how your assistant is set up.
• Context length at a glance: Each model in the dropdown shows its context window size, so you can make an informed choice without leaving the setup flow.

This update makes it much easier to find and use the right model for your assistant, whether you want the latest flagship or a cost-effective option for simpler tasks.

Hermes assistants now support inbound email. When someone sends a message to your assistant's email address, it arrives as a structured webhook event that the agent can process, respond to, and take action on. This brings Hermes to parity with OpenClaw on one of the most requested messaging channels.

• Email forwarding for Hermes: Inbound emails are routed to Hermes assistants through the same address-based forwarding that OpenClaw assistants already use, so setup works the same way.
• Structured webhook delivery: Hermes receives emails as typed webhook events with the sender, subject, body, and message threading information preserved, so the agent has full context to work with.
• Persistent storage for Hermes containers: Hermes assistants can now use durable storage that survives container restarts, so files, state, and workspace data stick around between sessions.
• More reliable terminal startup: The health sidecar now lazy-loads the terminal emulator, avoiding startup failures when the native module isn't immediately available inside the container.

Email is one of the simplest ways to interact with an assistant without installing anything. With this update, both OpenClaw and Hermes assistants can be reached the same way.

Hermes assistants now come with a built-in terminal in the dashboard. That means you can open a live shell directly inside the assistant's container, inspect files, run commands, and debug issues in place instead of guessing from the outside.

• A dedicated Terminal tab: Active Hermes assistants now have their own Terminal tab on the assistant details page, so shell access is part of the normal management flow instead of a separate workaround.
• Live shell access inside the container: You can connect to the running Hermes environment, execute commands, inspect the filesystem, and verify what is actually happening on the deployed assistant.
• Faster debugging when something feels off: If an assistant is behaving unexpectedly, you now have a direct way to check logs, files, processes, and runtime state without waiting for a redeploy or trying to infer everything from the UI.
• Secure, short-lived terminal sessions: Terminal access is gated behind signed session tickets and tied to the active assistant session, so it is designed for controlled troubleshooting rather than permanent open access.
• A more dependable Hermes runtime underneath: We also tightened the runtime path behind the scenes so terminal sessions connect more reliably and deploy status updates are less likely to fail during transient errors.

This is the most direct way we have added yet to work with a hosted assistant like a real running system, not just a dashboard object. For Hermes users especially, it makes day-to-day debugging and inspection much more practical.

getclaw now supports Hermes Agent as a hosted framework. That means you can create a Hermes assistant directly from the product, deploy it through the same managed flow, and operate it from the same dashboard you already use for OpenClaw assistants.

• A second hosted framework: You can now choose Hermes Agent instead of OpenClaw when creating a new assistant, directly inside getclaw.
• Managed through the same dashboard: Hermes assistants live inside the same deployment, management, and post-launch workflow as the rest of the platform, so there is no separate hosting path to learn.
• Telegram support for Hermes deployments: Hermes assistants can now run on Telegram in the hosted environment, so messaging works as part of the normal getclaw channel setup flow.
• Approval controls for Telegram access: When a restricted user messages a Hermes assistant, their Telegram user ID shows up in the dashboard so you can approve them without leaving getclaw.
• A more usable hosted Hermes runtime: Hermes deployments now boot cleanly in the managed environment, which makes hosted Hermes much closer to a real day-to-day option instead of an experimental add-on.

This is a framework expansion, not just a new toggle. getclaw now supports two different open-source agent runtimes, which gives users a clearer choice between OpenClaw's operations-first model and Hermes's more personal, persistent-agent model.

Until now, each conversation with your assistant started from a blank slate. The assistant could be brilliant in the moment, but the next time you talked to it, all of that context was gone. Lossless-claw changes that. It gives every managed assistant a persistent, structured memory that grows over time — so your assistant gets more useful the more you use it.

• Full conversation memory: Lossless-claw captures the complete context of every conversation — not just summaries or keywords, but the actual substance of what was discussed, decided, and learned.
• Automatic context recall: When your assistant starts a new conversation, it pulls in relevant context from past interactions. It knows what you've worked on, what you've asked before, and what preferences you've expressed — without you having to repeat yourself.
• Gets smarter over time: The more you use your assistant, the better it understands your projects, your terminology, and how you like things done. This is the difference between a tool that resets every session and one that actually learns.
• On by default, no setup: Every assistant now ships with lossless-claw pre-installed and active. There is nothing to configure. If you need to disable it for a specific assistant, set the DISABLE_LOSSLESS_CLAW environment variable.
• Baked into the image: The plugin is pre-installed in the container image at build time, so there is no download delay on cold starts and no risk of install failures at runtime.

This is a step change in what managed assistants can do. An assistant that remembers is fundamentally more useful than one that forgets — for ongoing projects, repeated workflows, and anything where continuity matters.

The infrastructure underneath managed assistants is now faster to start, easier to debug, and lighter on resources. This release removes unused dependencies from the container image, parallelizes data restores, and finally gives you visibility into what actually went wrong when a deploy fails.

• A smaller, faster container image: Playwright and pnpm have been removed from the Docker image. The result is a leaner image and faster startup.
• Parallel R2 restores: Workspace and skills data now restore from R2 concurrently instead of sequentially, cutting restore time on restarts.
• Delete backup data from the dashboard: You can now delete individual files or entire directories from your assistant's R2 backup storage directly in the dashboard, with streamed progress for large deletions.
• Safer Telegram defaults: If a Telegram allowlist is configured with no approved users, the runtime now falls back to pairing mode automatically instead of rejecting all messages.

Assistants start faster, and when something breaks during startup, you can actually see why.

Managing channels is now much cleaner, and the runtime underneath managed assistants just took a meaningful step forward too. This release adds a dedicated Messaging tab for channel setup and rolls hosted assistants onto OpenClaw 2026.3.13, which pulls in the biggest recent capability upgrades from upstream.

• A dedicated Messaging tab: Telegram, Slack, Discord, and Email now live in one place on the assistant details page, so channel setup no longer feels scattered across separate screens and controls.
• Easier channel management: You can enable or disable channels, rotate bot tokens, see whether a channel is configured, copy an assistant's email address, and jump straight into email management without leaving the dashboard.
• The big OpenClaw upgrade is now in managed deployments: Moving to OpenClaw 2026.3.13 brings the biggest recent upstream improvements into the hosted runtime, including browser automation, assistant email support, and unlimited built-in search.
• A safer and more predictable runtime: Deployments now validate the OpenClaw config before startup, set Control UI origins more explicitly for hosted environments, and pin Telegram streaming to a safer default so upgrades are less likely to break in surprising ways.

This is partly a dashboard release and partly an infrastructure one, but both sides matter. Channel operations are easier day to day, and managed assistants are now much closer to the current OpenClaw experience instead of lagging behind it.

Search is now available to every OpenClaw user, and it ships with unlimited usage from day one. No plan upgrade, extra toggle, or special setup required. If you already have assistants deployed, just redeploy each instance once to enable the new feature.

• Search is now built in: OpenClaw assistants can now use search as a native capability, opening up research and retrieval workflows that were not available before.
• Unlimited usage included: Every user gets the feature with no search cap, so assistants can keep using it without running into usage limits.
• One redeploy to enable it: Existing instances need a redeploy to pick up the new runtime behavior. Once redeployed, the feature is active automatically.
• Applies to your current assistants: This is not limited to newly created assistants. Your existing instances can use search too as soon as they are redeployed.

This is a meaningful capability release, not just a limit change. Search is now part of what every assistant can do, and if you already have one live, a redeploy is all it takes to turn it on.

OpenClaw assistants now have their own email address. They can check their inbox, send messages, and reply to threads — all as part of a workflow.

• Dedicated email address: Each assistant gets its own email address, provisioned automatically when the skill is enabled. No external setup required.
• Send, reply, and check inbox: The assistant can compose new emails, reply within existing threads, and read incoming messages on its own.
• Generous free tier: Every user gets 100 free emails per month. Beyond that, additional emails cost fractions of a cent from your credit balance.
• Threaded conversations: Replies are linked to the original message, so recipients see a clean conversation thread instead of disconnected messages.

Email turns your assistant from something you talk to into something that can talk to others on your behalf — following up, reporting, and communicating without you in the loop.

OpenClaw can now use a real browser as part of its workflow. That means your assistant is no longer limited to APIs and text-only tools. It can move through the web the way a person would: opening pages, clicking buttons, reading content, and acting on what it finds.

• Navigate real websites: OpenClaw can visit live pages, wait for interfaces to load, and work through multi-step browser flows instead of stopping at static HTML.
• Interact with the page: Clicking, form filling, scrolling, and page-to-page navigation are now part of the assistant's toolkit, which unlocks workflows that were previously impossible with API-only access.
• Read what users actually see: The assistant can inspect rendered pages, not just backend responses, which is a big step forward for research, operations, QA, and support-style tasks.
• Unlock broader automation: This creates room for assistants that verify dashboards, collect information from tools without public APIs, reproduce UI issues, and complete browser-based tasks end to end.

This is one of the biggest capability jumps in the product so far. Giving OpenClaw a browser turns it from an assistant that can call tools into one that can operate directly on the web.

This release is about polish where it matters: assistants start more smoothly, the first-run experience feels more alive, and Telegram access is less likely to surprise you later.

• More dependable launches: Deployments now recover more gracefully during slower startup moments, so assistants are less likely to get stuck while going live.
• Telegram access that stays put: If you maintain a private Telegram allowlist, those approved users now stay approved more reliably across edits and redeploys.
• Clearer save and redeploy feedback: When you update files from the dashboard, you now get much better progress messaging so it's obvious what was saved and what's still rolling out.
• A more premium first impression: Persona previews load faster, hover demos feel smoother, and the homepage greeting feels more intentional instead of generic.

Small release on paper, but it removes a lot of friction in the moments users actually notice: launch, preview, and trust.

Creating an assistant no longer means filling out forms. Just describe what you want in a conversation, and the wizard configures everything for you.

• Chat-based creation: Describe your assistant in plain language — name, personality, model, API key, and channels are all set through a guided conversation.
• 6 pre-built personas: Get started instantly with LinkedIn Influencer, News Reporter, Sales Rep, Customer Support, Software Developer, or Personal Coach — each with a ready-to-go personality and skills.
• Redesigned homepage: A new homepage showcases what's possible with persona demos and video walkthroughs.
• Fresh visual identity: Updated logo, favicons, and branding across the product.

Two improvements to help you understand and manage running assistants faster.

• Live server logs: A new logs panel on the assistant detail page streams your worker's output in real time, with color-coded log levels and auto-scrolling.
• Playback controls: Start, stop, and clear the log stream without leaving the dashboard.
• Faster config changes: Updating your model or API key now takes effect in about 30 seconds, down from a full redeploy that could take several minutes.
• Visual update status: The dashboard shows an "Updating configuration..." indicator while changes are being applied.

Your assistant is no longer limited to a single messaging platform. You can now deploy to Telegram, Slack, Discord, or any combination of the three.

• Slack support: Connect your assistant to a Slack workspace using a Bot Token and App Token. A built-in setup guide walks you through creating the Slack app.
• Discord support: Connect your assistant to a Discord server with a single bot token. The setup guide covers bot creation, permissions, and server invitation.
• Multi-channel deployments: Enable multiple channels on a single assistant. The detail page shows all active channels at a glance.
• Step-by-step setup guides: Each channel includes an expandable guide with the exact steps, scopes, and permissions needed.

Until now, editing assistant files meant working through openclaw UI or redeploying from scratch. The new file editor gives you direct access to your assistant's files right from the dashboard.

• Browse your assistant's files: Navigate directories, open files, and explore your assistant's configuration and workspace from a familiar file explorer.
• Edit with syntax highlighting: A full-featured code editor with automatic language detection for JSON, TypeScript, JavaScript, YAML, Markdown, shell scripts, and more.
• Save & redeploy in one click: Edit a file and redeploy your assistant in a single action, no need to leave the dashboard.
• Safe by default: Unsaved changes are flagged before closing, and saves are validated to prevent conflicts.

This update improves how managed assistants preserve state, so restarts are less disruptive and recovery is more dependable.

• Better data safety: Backups now run more consistently, including periodic full backups to reduce the chance of drift over time.
• Smoother restarts: Assistants now recover newer saved state more reliably when containers restart.
• More complete recovery: Restores now cover assistant state, workspace files, and installed skills.
• Clear backup health: Admin status now shows last successful backup, last backup attempt, and recent backup errors when present.
• Customer value: Less manual recovery work, fewer surprises after restarts, and more confidence that assistant progress is preserved.

We made this change because cloud browser automation setups were too fragile and required too much manual configuration. The goal of this release is to give you safer defaults, fewer deployment steps, and more reliable assistant behavior in isolated environments.

• Why this matters: Cloud environments can fail in ways that are hard to debug when browser connectivity is not configured securely and consistently.
• What you now get: Deployments now apply secure connectivity defaults automatically, so assistants come up with fewer manual steps.
• Reliability improvement: Isolated sandboxes now default to the Playwright CLI browser path, which avoids common extension-relay failures.
• Stability improvement: Gateway configuration is normalized for current OpenClaw runtimes, reducing startup and connectivity issues.
• Customer value: Faster time-to-working assistant, less deployment troubleshooting, and safer browser automation defaults out of the box.

We are introducing first-class support for fully customizing your openclaw.json, alongside making assistant configuration easier to understand and more dependable.

• Customize openclaw.json: You now have direct control to customize your openclaw.json configuration just how you want it.
• Clearer config status: You can now quickly see which settings are currently applied versus still pending.
• Easier troubleshooting: Side-by-side config comparisons make it faster to spot what changed.
• More reliable updates: Config saves are now applied more consistently across assistant redeploys.
• Cleaner comparisons: Runtime-managed values are filtered out so you can focus on user-controlled changes.
• Supported OpenClaw version: 2026.2.19.

This release focused on product clarity and assistant configuration flow.

• Model updates are easier: Change your assistant model directly from the assistant details page.
• Cleaner logged-in experience: Returning users are now guided to their assistants right away.
• Improved onboarding content: New comparison content helps users choose the right path faster.

This update focused on clarity and usability.

• Smoother account flows: Improved key flows so setup and activation feel more reliable.
• Better docs navigation: Updated structure and links so it is easier to find what you need.
• Clearer guides: Expanded and refined help content for common usage patterns.

This release made external integrations more complete while improving day-to-day stability.

• More complete integrations: Added broader support for assistant management tasks.
• Easier setup: Updated guides and behavior for a smoother integration experience.
• Stability improvements: Fixed issues that could affect account, billing, and assistant operations.

This patch focused on making managed usage more dependable.

• More accurate credits: Resolved edge cases around balance updates and top-ups.
• More stable assistant performance: Fixed issues that could interrupt managed assistant responses.

A full pass on the landing experience improved readability and conversion-oriented messaging.

• New page structure: Improved layout and section flow to make the value proposition clearer.
• Visual and copy polish: Updated visuals and messaging to improve readability and conversion.

This was a major milestone for managed assistant billing and reliability.

• Managed credits are live: Users can now run assistants with platform-managed credits.
• Improved account isolation: Credits are now separated per user for better consistency.
• Updated pricing structure: Managed assistant plans were clarified and separated from other offerings.
• Deployment reliability improvements: Managed assistant deployment flow is now more dependable.

Payments and production reliability moved from setup to functional rollout.

• Payments are live: Added paid plans and credit purchase flows for users.
• Sign-in fix: Resolved a Google sign-in issue that affected some production users.
• Reliability improvements: Fixed several production issues to improve overall stability.
• Legacy branding cleanup: Removed lingering legacy naming in UI/docs and standardized branding.

This release focused on improving security for critical user actions.

• Stronger protections: Added additional safeguards around sensitive platform operations.
• Safer event handling: Improved how background events are validated and processed.
• Improved request checks: Tightened validation for account and assistant-related actions.

This is the first public foundation of the product.

• Core product launch: Released assistants dashboard, authentication, and essential user flows.
• Assistant lifecycle support: Enabled creating, managing, and operating assistants end to end.
• Foundational docs and tools: Shipped initial docs and setup tooling.

Shared Hermes chat links now give people more than a single message thread. The chat page has a sidebar for switching conversations, seeing what the assistant can do, and reaching owner settings when you have access.

• Chat sessions in the sidebar: Start a fresh chat, return to recent conversations, and keep each shared link organized on the device where it is being used.
• Skills are visible from chat: The Skills view shows the assistant's installed capabilities with descriptions, so users can understand what the assistant is equipped to handle before they ask.
• Owner controls stay close: Assistant owners can jump from the shared chat to Secrets, while public visitors only see the controls that make sense for them.
• Cleaner shared chat UI: The chat header moved into the sidebar, theme controls are easier to find, and the page keeps more room focused on the conversation.
• More dependable responses: Slow or interrupted Hermes replies now fail more clearly instead of leaving the chat stuck waiting.

This makes shared Hermes links easier to hand to customers, teammates, and collaborators. They can chat, understand the assistant's skills, and keep separate sessions without needing the full owner dashboard.

WhatsApp assistants now have more precise reply rules. You can keep one default instruction for everyone, then add overrides for the exact people or groups where Hermes should behave differently.

• Everyone, numbers, and groups: The Rules section now separates the global default from specific contacts and WhatsApp groups, so one assistant can handle different conversations with different guidance.
• Contact-specific instructions: Add a number or pick a WhatsApp contact, then tell Hermes how to answer that person. That means VIP clients, suppliers, or teammates can get tailored behavior without changing the default rule.
• Group-specific instructions: Give busy groups their own policy, like answering only direct requests for help and ignoring casual chatter.
• WhatsApp search in setup: Search contacts and groups from the Rules panel instead of copying chat ids by hand. Recent inbox chats still help when a live address book is warming up.

This makes WhatsApp automation easier to use in real operations. Hermes can stay broad by default, then get stricter, quieter, or more specific exactly where the conversation needs it.

WhatsApp auto-reply is getting safer for real customer conversations. You can now define what Hermes must not disclose, tune how much of its work is visible in chat, and manage pairing and automation from a clearer WhatsApp setup flow.

• Disclosure guardrails: Add rules for private details Hermes should never reveal, like internal supplier names, payment details, customer phone numbers, private calendar notes, or unpublished discounts.
• Reviewed replies before delivery: When guardrails are active, drafted WhatsApp replies are checked before they are sent. If a draft crosses the line, Hermes rewrites it or falls back to a safe refusal.
• Cleaner customer-facing messages: Turn off the visible [hermes] marker when you want auto-replies to read like normal WhatsApp messages.
• Optional thinking steps: Choose whether contacts see Hermes tool-progress bubbles while it works, or only the final answer.
• Clearer WhatsApp setup: Pairing, message output, reply scope, instructions, allowlists, and guardrails now live in a more structured WhatsApp panel, so it is easier to see what is ready and what still needs attention.

This makes WhatsApp assistants easier to use in higher-trust conversations. You can keep Hermes helpful while controlling what it says, how it appears, and when it stays quiet.

Some hosted apps are meant for a specific client or team. Others are meant to be opened by anyone with the link. You can now mark a Hermes-hosted app as public directly from the Apps tab, while keeping the rest of your assistant's apps private or limited to the exact people you choose.

• Public hosted apps: Use the new public option when a dashboard, report, website, or demo should open without sign-in. That means customer-facing work can be shared as a normal link.
• Private apps stay protected: Owner-only, email-specific, and company-domain access still work for the apps that should stay internal.
• Clearer access state: App cards now show when an app is public, and reopening the access panel keeps that state visible instead of making public apps look owner-only.
• Safer public switching: Public access is synced with the edge protection in front of the app and with the assistant's own app registry, so the link behavior and dashboard state stay aligned.
• Healthier Hermes gateways: Active Hermes assistants are checked in the background, and unhealthy gateways are repaired automatically when the assistant is not already deploying or updating.

This makes hosted Hermes apps easier to use as real deliverables. You can publish something private, share it with a narrow audience, or make it fully public without changing how the assistant itself is managed.

You can now decide who gets into each hosted app your Hermes assistant publishes. A private dashboard can stay owner-only, while a client report or team page can be opened to the exact emails or domains you choose.

• Per-app sharing controls: Every hosted app gets its own Access panel in the Apps tab, so one dashboard can be shared with a client while another stays private.
• Invite by email: Add individual people when you want a small, exact audience for a dashboard, report, or internal tool.
• Allow a whole domain: Share with everyone at a company by adding a domain like example.com, without typing every teammate by hand.
• Owner access stays protected: The assistant owner always keeps access, even when the shared audience changes.
• Clear visibility in the Apps tab: App cards now show whether an app is owner-only or shared, so it's easier to audit what is exposed.

This turns hosted Hermes apps into something you can confidently share. You still get protected links, but now the audience can match the work instead of being all-or-nothing.

Hermes assistants can now turn generated work into live dashboards and websites. Instead of stopping at files in a session, your assistant can publish a protected app that stays attached to the assistant and shows up in GetClaw.

• Publish from the assistant: Ask Hermes to create or update a dashboard, report, landing page, or small website and it can make that work available through the assistant's own live link.
• New Apps tab: Every Hermes assistant now has an Apps tab that lists published surfaces with names, descriptions, update times, and a direct Open button.
• Protected by default: Published apps use the same sign-in protection as the assistant, so private dashboards and internal pages do not become public just because they are easy to open.
• No extra hosting setup: Hermes can publish the finished work itself. You do not need to connect Vercel, Netlify, Cloudflare, or a separate server just to view the result.
• Built for living dashboards: Published apps can be refreshed as the assistant gathers new information, so a report can become a reusable dashboard instead of a one-off artifact.

This makes Hermes more useful after it finishes the first task. The output can become a place you return to, share with the right people, and keep improving with the same assistant.

WhatsApp auto-reply is now more useful in real conversations. Instead of relying on a fixed trigger word or answering every eligible chat the same way, you can write plain-language instructions that tell Hermes when it should answer and when it should stay quiet.

• Custom auto-reply instructions: Add instructions like “answer only when someone is trying to book a dining slot” directly from the WhatsApp panel. Hermes uses that rule before sending anything back.
• Quiet by design: If a message does not match your instructions, Hermes sends no WhatsApp reply. That means you can connect busier chats without creating noise.
• All DMs + groups is smarter: The group-chat mode now evaluates inbound DMs and group messages with your instructions instead of requiring people to start every message with a special prefix.
• Default modes still work: Self and allowlist modes keep their familiar behavior. Leave the instructions blank when you want the normal generic auto-reply, or add guidance when you want more control.
• Visible typing feedback: WhatsApp now shows that your assistant is typing while Hermes is working, then clears the indicator when the reply is done or when Hermes decides not to answer.

This makes WhatsApp feel less like a bot switch and more like a real assistant policy. You decide the situations where Hermes should participate, and the assistant stays out of everything else.

Until today, your WhatsApp assistant only worked in direct messages, and the “All DMs” mode answered every single incoming chat. Useful for solo testing, noisy in real life. Group chats were off-limits entirely. Both of those things change with a small new convention: the word gg.

• Group chats are in: Add your assistant's WhatsApp account to any group and it'll reply on demand. It stays quiet unless someone explicitly calls it.
• gg is the trigger: Type gg followed by your question, like gg what's on my calendar today?, and only that message reaches Hermes. Anything without the prefix is ignored, so normal group chatter doesn't spawn replies.
• Call it from your own phone too: When you're sitting in a group on the phone your assistant is paired to, type gg ... yourself and the reply lands right in the group. Great for asking your assistant in front of the people you're planning with.
• Case-insensitive, prefix stripped: gg, GG, and Gg all work. The prefix is removed before Hermes sees the message, so your prompts stay clean and the answers don't echo the trigger word back.
• Spelled out in the WhatsApp panel: The auto-reply mode is now called All DMs + groups. Pick it on any WhatsApp-connected assistant and an inline explainer shows the trigger word, a working example, and exactly which messages will be ignored.
• Self and Allowlist modes are unchanged: If you'd rather keep your assistant scoped to your own “Me” chat or a hand-picked list of contacts, those modes still work the way they did, with no prefix needed and groups still off.

Groups have been on the wishlist since the day we shipped WhatsApp. Now you can drop your assistant into a planning thread, a family chat, or a small team group, and pull it into the conversation only when you actually want it to chime in.

Until now, talking to your Hermes assistant meant going through a chat channel like WhatsApp, Telegram, or iMessage. There was no clean way to put it inside your own product. A new public Chat API changes that. Every Hermes assistant gets its own OpenAI-compatible endpoint, scoped to that assistant, that you can call from any app you build.

• Drop-in for any OpenAI SDK: The endpoint speaks the same shape as OpenAI's chat completions API. Set your getclaw API key as the SDK's key, point the base URL at your assistant, and the rest of your code stays exactly the way you wrote it.
• Stream replies token by token: Responses come back over a streaming connection, so your UI can render the reply as it's being generated. Cancel from your side and the assistant stops working on it too, no wasted tokens.
• Authenticated with your getclaw API key: The same key you already use to manage assistants now unlocks the Chat API. Send it as a Bearer token or an X-API-Key header. No separate keys to juggle.
• Server-side memory per end user: Pass an X-Hermes-Session-Id header and your assistant keeps the conversation going across calls, on its side, scoped to that id. Skip the header for stateless one-shot calls. Either pattern works without you running a database.
• A new Chat API tab on every Hermes assistant: Open any Hermes assistant in the dashboard and you'll find a Chat API tab with copy-paste curl, OpenAI SDK, and session-aware snippets pre-filled with your deployment. Grab a working request in seconds.
• Documented in the API reference: New entries for chat completions and the model list show up in the API reference and the OpenAPI spec, so the same docs your team already reads now describe the chat surface too.

The whole point of giving your assistant skills, memory, and channels is to make it useful in the places your users already are. Now one of those places can be your own product, behind your own UI, with the assistant doing the work.

If you'd rather not register a number with a third-party iMessage provider, you can now point Hermes at a BlueBubbles server running on a Mac you own. Replies go out from your real iMessage account, on the device you already use, with nothing else in the middle.

• Use your own iMessage account: Replies come from the Apple ID signed into your Mac. No new number, no shared shortcode, no extra middleman holding your conversations.
• One panel, one webhook: Open the iMessage via BlueBubbles tile on any Hermes assistant, paste your server URL and password, and copy the webhook URL we generate straight into BlueBubbles. The setup guide walks you through every step.
• You decide who it answers: Start with an allowlist of specific chats and grow it over time, or open it to every contact at once. Anyone outside the list gets nothing back.
• Pause without unpairing: A single switch on the panel turns auto-reply off while keeping your connection live. Flip it back on whenever you're ready and your assistant picks right up.
• Server password stays encrypted: We store your BlueBubbles password encrypted and never echo it back to the page. Rotate it any time by entering a new one, and the same webhook URL keeps working.
• Disconnect anytime: One click clears the credentials, stops your assistant from answering, and leaves your Mac and your threads exactly where they were.

For people who already keep a Mac running at home or in the office, this is the most direct way to put an assistant on iMessage: your account, your hardware, your control, with the convenience of a managed setup flow on top.

Until now, the only env vars your assistant could see were the ones we set up ourselves. Bringing in a Stripe key, a Notion token, or any other API credential meant editing files inside the container by hand and crossing your fingers it survived the next redeploy. A new Secrets tab on every Hermes assistant fixes that.

• See what's already injected: The Platform-managed list shows every variable we set on your behalf, with a short note on what each one does. Empty slots and internal plumbing are hidden, so you only see things that actually matter for your assistant.
• Reveal any value on click: Values are masked by default. Hit Show on any row to see the real key, token, or URL, and it auto-hides again after a few seconds so it doesn't sit on your screen.
• Add your own keys for skills: Drop a custom variable into the Your Variables list and your skills can read it as a normal environment variable from inside the container. Names like STRIPE_API_KEY or NOTION_TOKEN are yours to define.
• Stack changes, redeploy once: Click Add Variable as many times as you need; each click drops a new draft row in the table. Fill them all in, then hit Redeploy once and they all go live together. No more one-redeploy-per-key.
• Edit and remove without leaving the page: Update a value or pull a custom variable straight from the table. Changes batch up alongside your drafts so you only push to the live assistant when you're ready.
• Encrypted at rest, persistent across redeploys: Anything you add is encrypted in your account and re-injected on every boot. Custom variables stick around through redeploys, region moves, and model changes without you having to re-enter them.
• Safe by default: Names that collide with the variables we manage for you are rejected up front, so you can't accidentally clobber your provider key or a channel token by typing the wrong thing.

Most of what makes an assistant useful in real life is the services it can reach. Now you can hand yours any credential it needs, see exactly what it's working with, and change your mind whenever, all from one tab.

iMessage is now a channel for Hermes assistants, powered by Sendblue. Register a number with Sendblue, paste the credentials into the dashboard, point Sendblue at the webhook we generate, and inbound iMessages start flowing to your assistant. Replies go back through the same number, with the blue bubble.

• Set it up in one panel: Open the iMessage tab on any Hermes assistant, drop in your Sendblue API key, secret, and registered number, and copy the webhook URL and secret straight into Sendblue. The whole loop fits on one screen with a step-by-step guide right there.
• Bring your own number: The number that texts your contacts is the one you registered with Sendblue, not a shared shortcode or a stranger's line. People see a normal blue bubble from a number you control.
• You decide who it answers: Start with an allowlist and add the specific chats your assistant should reply to. Open it up to everyone in your contacts later, or keep it tight forever. Anyone outside the list gets nothing.
• Toggle auto-reply without unpairing: One switch on the iMessage panel pauses replies while keeping the connection live. Flip it back when you're ready and your assistant picks right up.
• Credentials stay encrypted: Your Sendblue API secret is stored encrypted and never echoed back to the page. Rotate it whenever you want by entering a new one and saving.
• Disconnect anytime: One click clears the credentials and stops the assistant from answering. Sendblue keeps your number, your contacts keep their threads, and nothing else changes.

iMessage is where a lot of personal and small-team conversations actually happen. Putting your assistant in that thread, behind a number you own and a list you control, makes it reachable in the place people already text from.

The interactive shell on every Hermes assistant page has a new emulator under the hood. It's the same engine behind Ghostty, the modern terminal, ported to run directly in your browser. Same Connect, Reconnect, and Clear buttons you already use, with a much better experience once you're typing.

• Powered by Ghostty: The emulator now handles escape sequences, Unicode, and color codes the way a serious native terminal does. Output that used to come back garbled or half-cleared now lands cleanly.
• TUI apps just work: Run vim, htop, nano, less, or any ncurses-based tool inside your container and it renders correctly, redraws cleanly, and gives back a tidy prompt when you exit.
• Real text selection and copy: Drag to select, double-click to grab a word, Cmd+C to copy. The terminal renders to the page like ordinary text, so the selection behavior you already know works here too.
• Smoother resizing: Resize your browser window and the terminal reflows immediately, telling the running shell its new dimensions so prompts and TUIs stay aligned without a manual redraw.
• Same controls, no migration: Connect, Reconnect, Clear, and Disconnect all live exactly where they did. The change is entirely under the hood — open a session and you'll feel the difference, not see a new UI.

The container terminal is the place you reach for when something isn't behaving and you want to look at it directly. Making it accurate, native to the page, and out of your way matters more than it gets credit for.

WhatsApp is now a first-class channel for Hermes assistants. Pair it once from the dashboard and the assistant lives inside your existing WhatsApp account, replying to the people and threads you trust without anyone needing to install anything new.

• Pair with a QR code: Open the WhatsApp tab on any Hermes assistant and scan the QR with your phone, the exact flow you already use for WhatsApp Web. The session stays linked until you unpair it, so you only do this once.
• No Business account, no extra number: Your assistant uses your real WhatsApp account. There's nothing to register with Meta, no separate SIM to buy, and no approval queue to sit in before you can talk to anyone.
• You decide who it answers: An allowlist on the assistant page controls exactly which contacts and groups the assistant replies to. Everyone else sees nothing different, so your normal chats stay normal.
• Handles photos, voice notes, and documents: Send your assistant an image, a voice memo, or a PDF and it can read, transcribe, or work on it just like a text message. It can also send media back, so replies aren't stuck in plain text.
• Talks to itself when you want it to: Message your own number and the assistant treats it as a private command line. Great for quick captures, reminders, and one-off tasks without spinning up a separate chat.
• Unpair anytime: One click on the dashboard ends the session and removes the assistant from your WhatsApp account. Your messages and contacts stay where they always were.

WhatsApp is where most of the world already messages. Putting your assistant there means people can reach it without learning a new app, downloading anything, or changing how they talk to you.

Until now, every assistant ran in whichever region happened to be closest to wherever it first spun up. You can now choose the region yourself from the assistant's configuration, with nine options spanning the major continents.

• Pick from nine regions: Eastern and Western North America, Western and Eastern Europe, Asia Pacific, South America, the Middle East, Oceania, and Africa. The picker calls out regions with limited capacity so you know what to expect.
• Defaults to Eastern North America: New assistants come up in Eastern North America by default. If that's where most of your users or data already live, you don't need to do anything.
• Change it whenever you want: Open your assistant's details page, hit the pencil next to Region, pick a new one, and save. If your assistant is live, a redeploy starts automatically so the new region takes effect.
• Better latency for non-US users: If your assistant chats mostly with people in Europe or Asia, moving it closer to them shortens the round trip on every message and tool call.
• Keeps traffic close to your data: Calling APIs or databases hosted in a particular region? Run your assistant in the same one and you save a hop on every request.

Latency adds up over a long conversation, especially when tools are involved. Putting the assistant in the right place is one of the easiest ways to make it feel snappier without changing anything else.

Every assistant inbox now has a handle you control. Instead of the auto-generated address you got when email was first enabled, you can pick something short and memorable and hand it out with confidence. The change applies to both OpenClaw and Hermes assistants.

• Choose your own handle: On the Email tab for any assistant, type the name you want before the @ and save. Your inbox address updates immediately and any future email sent by the assistant goes out from the new address.
• Change it whenever you want: Renamed your assistant, or just want something catchier? Update the handle again and the new address takes effect right away. No redeploy to babysit on OpenClaw, and a seamless redeploy on Hermes.
• Globally unique, guaranteed: Handles are reserved the instant you claim one, so two assistants can never share an inbox. If someone else has already taken the name you want, you'll see a clear message and can try another.
• Friendly validation as you type: The form tells you right away if a handle is too short, uses unsupported characters, or touches one of the reserved mailbox names. No guessing, no trial-and-error saves.
• Defaults unchanged if you don't care: Assistants that already have email enabled keep their existing address — nothing moves until you choose a new handle.
• A cleaner delete confirmation: Deleting an assistant now opens a proper in-app dialog instead of a browser prompt, with a loading state while the tear-down runs and a smooth exit animation on the assistants list.

A random identifier is fine for an internal ID, but it's a terrible thing to hand out to colleagues, customers, or vendors. Now your assistant's inbox can look like it belongs to a person, not a database row.

Hermes is now more than 50% faster. Below is time to first reply — how long between a user's message and the assistant's first word back — tracked over the last few weeks. The baseline held steady for a long time; this release is the cliff on the right.

Two shifts landed in this release. Replies are dramatically faster (above), and assistants now keep their full state across redeploys — every conversation, uploaded file, scheduled job, and approved pairing survives.

• Over 50% faster replies: Simple questions that used to take about 30 seconds now come back in around 7. Complex tasks that took a full minute now finish in roughly 20.
• Automatic snapshots on redeploy: The moment you trigger a redeploy, your assistant's state is captured and saved before the running instance is replaced. Nothing extra to configure, nothing to remember.
• Seamless restore on the new instance: When the replacement instance comes up, it boots with the exact workspace, memory, and channel pairings the previous one had. Conversations pick up right where they left off.
• Existing assistants migrated automatically: Hermes assistants created before this change have their prior data moved over into the new snapshot format, so long-running assistants keep their history without any manual work.
• More forgiving redeploys: A redeploy no longer fails just because your assistant is busy working through a backlog or waiting on an external provider. The platform gives it room to settle instead of restart-looping.
• Faster channel reconnection: After a redeploy, Telegram pairings and other approved channels are already restored, so there's no re-pairing or re-authorization step.

Redeploys used to be the scariest button in the product for long-running Hermes assistants, and waiting on replies was the slowest part of a typical day. With this release, shipping an update is routine instead of a data-loss risk, and your assistant answers while you're still looking at the screen.

When you create a managed assistant, the model picker now pulls in the full OpenRouter catalog instead of a short curated list. That means you can browse and search across 500+ models from dozens of providers, all in the same dropdown. We also updated the defaults so new assistants start on the latest models.

• Full model catalog: The model selector during onboarding now shows every model available through OpenRouter, organized by provider. Popular providers like Anthropic, OpenAI, Google, and DeepSeek appear first, and you can type to search the rest.
• Instant search across all models: Start typing a model name, provider, or ID and the list filters down immediately. No more scrolling through a fixed list hoping the model you want is there.
• GPT-5.4 Mini as the new default: Managed assistants now default to OpenAI's GPT-5.4 Mini, which offers strong performance at a lower price point. You can always change it before confirming.
• GPT-5.4 and GPT-5.4 Mini added: Both models are now available for managed and bring-your-own-key assistants, so you can use them regardless of how your assistant is set up.
• Context length at a glance: Each model in the dropdown shows its context window size, so you can make an informed choice without leaving the setup flow.

This update makes it much easier to find and use the right model for your assistant, whether you want the latest flagship or a cost-effective option for simpler tasks.

Hermes assistants now support inbound email. When someone sends a message to your assistant's email address, it arrives as a structured webhook event that the agent can process, respond to, and take action on. This brings Hermes to parity with OpenClaw on one of the most requested messaging channels.

• Email forwarding for Hermes: Inbound emails are routed to Hermes assistants through the same address-based forwarding that OpenClaw assistants already use, so setup works the same way.
• Structured webhook delivery: Hermes receives emails as typed webhook events with the sender, subject, body, and message threading information preserved, so the agent has full context to work with.
• Persistent storage for Hermes containers: Hermes assistants can now use durable storage that survives container restarts, so files, state, and workspace data stick around between sessions.
• More reliable terminal startup: The health sidecar now lazy-loads the terminal emulator, avoiding startup failures when the native module isn't immediately available inside the container.

Email is one of the simplest ways to interact with an assistant without installing anything. With this update, both OpenClaw and Hermes assistants can be reached the same way.

Hermes assistants now come with a built-in terminal in the dashboard. That means you can open a live shell directly inside the assistant's container, inspect files, run commands, and debug issues in place instead of guessing from the outside.

• A dedicated Terminal tab: Active Hermes assistants now have their own Terminal tab on the assistant details page, so shell access is part of the normal management flow instead of a separate workaround.
• Live shell access inside the container: You can connect to the running Hermes environment, execute commands, inspect the filesystem, and verify what is actually happening on the deployed assistant.
• Faster debugging when something feels off: If an assistant is behaving unexpectedly, you now have a direct way to check logs, files, processes, and runtime state without waiting for a redeploy or trying to infer everything from the UI.
• Secure, short-lived terminal sessions: Terminal access is gated behind signed session tickets and tied to the active assistant session, so it is designed for controlled troubleshooting rather than permanent open access.
• A more dependable Hermes runtime underneath: We also tightened the runtime path behind the scenes so terminal sessions connect more reliably and deploy status updates are less likely to fail during transient errors.

This is the most direct way we have added yet to work with a hosted assistant like a real running system, not just a dashboard object. For Hermes users especially, it makes day-to-day debugging and inspection much more practical.

getclaw now supports Hermes Agent as a hosted framework. That means you can create a Hermes assistant directly from the product, deploy it through the same managed flow, and operate it from the same dashboard you already use for OpenClaw assistants.

• A second hosted framework: You can now choose Hermes Agent instead of OpenClaw when creating a new assistant, directly inside getclaw.
• Managed through the same dashboard: Hermes assistants live inside the same deployment, management, and post-launch workflow as the rest of the platform, so there is no separate hosting path to learn.
• Telegram support for Hermes deployments: Hermes assistants can now run on Telegram in the hosted environment, so messaging works as part of the normal getclaw channel setup flow.
• Approval controls for Telegram access: When a restricted user messages a Hermes assistant, their Telegram user ID shows up in the dashboard so you can approve them without leaving getclaw.
• A more usable hosted Hermes runtime: Hermes deployments now boot cleanly in the managed environment, which makes hosted Hermes much closer to a real day-to-day option instead of an experimental add-on.

This is a framework expansion, not just a new toggle. getclaw now supports two different open-source agent runtimes, which gives users a clearer choice between OpenClaw's operations-first model and Hermes's more personal, persistent-agent model.

Until now, each conversation with your assistant started from a blank slate. The assistant could be brilliant in the moment, but the next time you talked to it, all of that context was gone. Lossless-claw changes that. It gives every managed assistant a persistent, structured memory that grows over time — so your assistant gets more useful the more you use it.

• Full conversation memory: Lossless-claw captures the complete context of every conversation — not just summaries or keywords, but the actual substance of what was discussed, decided, and learned.
• Automatic context recall: When your assistant starts a new conversation, it pulls in relevant context from past interactions. It knows what you've worked on, what you've asked before, and what preferences you've expressed — without you having to repeat yourself.
• Gets smarter over time: The more you use your assistant, the better it understands your projects, your terminology, and how you like things done. This is the difference between a tool that resets every session and one that actually learns.
• On by default, no setup: Every assistant now ships with lossless-claw pre-installed and active. There is nothing to configure. If you need to disable it for a specific assistant, set the DISABLE_LOSSLESS_CLAW environment variable.
• Baked into the image: The plugin is pre-installed in the container image at build time, so there is no download delay on cold starts and no risk of install failures at runtime.

This is a step change in what managed assistants can do. An assistant that remembers is fundamentally more useful than one that forgets — for ongoing projects, repeated workflows, and anything where continuity matters.

The infrastructure underneath managed assistants is now faster to start, easier to debug, and lighter on resources. This release removes unused dependencies from the container image, parallelizes data restores, and finally gives you visibility into what actually went wrong when a deploy fails.

• A smaller, faster container image: Playwright and pnpm have been removed from the Docker image. The result is a leaner image and faster startup.
• Parallel R2 restores: Workspace and skills data now restore from R2 concurrently instead of sequentially, cutting restore time on restarts.
• Delete backup data from the dashboard: You can now delete individual files or entire directories from your assistant's R2 backup storage directly in the dashboard, with streamed progress for large deletions.
• Safer Telegram defaults: If a Telegram allowlist is configured with no approved users, the runtime now falls back to pairing mode automatically instead of rejecting all messages.

Assistants start faster, and when something breaks during startup, you can actually see why.

Managing channels is now much cleaner, and the runtime underneath managed assistants just took a meaningful step forward too. This release adds a dedicated Messaging tab for channel setup and rolls hosted assistants onto OpenClaw 2026.3.13, which pulls in the biggest recent capability upgrades from upstream.

• A dedicated Messaging tab: Telegram, Slack, Discord, and Email now live in one place on the assistant details page, so channel setup no longer feels scattered across separate screens and controls.
• Easier channel management: You can enable or disable channels, rotate bot tokens, see whether a channel is configured, copy an assistant's email address, and jump straight into email management without leaving the dashboard.
• The big OpenClaw upgrade is now in managed deployments: Moving to OpenClaw 2026.3.13 brings the biggest recent upstream improvements into the hosted runtime, including browser automation, assistant email support, and unlimited built-in search.
• A safer and more predictable runtime: Deployments now validate the OpenClaw config before startup, set Control UI origins more explicitly for hosted environments, and pin Telegram streaming to a safer default so upgrades are less likely to break in surprising ways.

This is partly a dashboard release and partly an infrastructure one, but both sides matter. Channel operations are easier day to day, and managed assistants are now much closer to the current OpenClaw experience instead of lagging behind it.

Search is now available to every OpenClaw user, and it ships with unlimited usage from day one. No plan upgrade, extra toggle, or special setup required. If you already have assistants deployed, just redeploy each instance once to enable the new feature.

• Search is now built in: OpenClaw assistants can now use search as a native capability, opening up research and retrieval workflows that were not available before.
• Unlimited usage included: Every user gets the feature with no search cap, so assistants can keep using it without running into usage limits.
• One redeploy to enable it: Existing instances need a redeploy to pick up the new runtime behavior. Once redeployed, the feature is active automatically.
• Applies to your current assistants: This is not limited to newly created assistants. Your existing instances can use search too as soon as they are redeployed.

This is a meaningful capability release, not just a limit change. Search is now part of what every assistant can do, and if you already have one live, a redeploy is all it takes to turn it on.

OpenClaw assistants now have their own email address. They can check their inbox, send messages, and reply to threads — all as part of a workflow.

• Dedicated email address: Each assistant gets its own email address, provisioned automatically when the skill is enabled. No external setup required.
• Send, reply, and check inbox: The assistant can compose new emails, reply within existing threads, and read incoming messages on its own.
• Generous free tier: Every user gets 100 free emails per month. Beyond that, additional emails cost fractions of a cent from your credit balance.
• Threaded conversations: Replies are linked to the original message, so recipients see a clean conversation thread instead of disconnected messages.

Email turns your assistant from something you talk to into something that can talk to others on your behalf — following up, reporting, and communicating without you in the loop.

OpenClaw can now use a real browser as part of its workflow. That means your assistant is no longer limited to APIs and text-only tools. It can move through the web the way a person would: opening pages, clicking buttons, reading content, and acting on what it finds.

• Navigate real websites: OpenClaw can visit live pages, wait for interfaces to load, and work through multi-step browser flows instead of stopping at static HTML.
• Interact with the page: Clicking, form filling, scrolling, and page-to-page navigation are now part of the assistant's toolkit, which unlocks workflows that were previously impossible with API-only access.
• Read what users actually see: The assistant can inspect rendered pages, not just backend responses, which is a big step forward for research, operations, QA, and support-style tasks.
• Unlock broader automation: This creates room for assistants that verify dashboards, collect information from tools without public APIs, reproduce UI issues, and complete browser-based tasks end to end.

This is one of the biggest capability jumps in the product so far. Giving OpenClaw a browser turns it from an assistant that can call tools into one that can operate directly on the web.

This release is about polish where it matters: assistants start more smoothly, the first-run experience feels more alive, and Telegram access is less likely to surprise you later.

• More dependable launches: Deployments now recover more gracefully during slower startup moments, so assistants are less likely to get stuck while going live.
• Telegram access that stays put: If you maintain a private Telegram allowlist, those approved users now stay approved more reliably across edits and redeploys.
• Clearer save and redeploy feedback: When you update files from the dashboard, you now get much better progress messaging so it's obvious what was saved and what's still rolling out.
• A more premium first impression: Persona previews load faster, hover demos feel smoother, and the homepage greeting feels more intentional instead of generic.

Small release on paper, but it removes a lot of friction in the moments users actually notice: launch, preview, and trust.

Creating an assistant no longer means filling out forms. Just describe what you want in a conversation, and the wizard configures everything for you.

• Chat-based creation: Describe your assistant in plain language — name, personality, model, API key, and channels are all set through a guided conversation.
• 6 pre-built personas: Get started instantly with LinkedIn Influencer, News Reporter, Sales Rep, Customer Support, Software Developer, or Personal Coach — each with a ready-to-go personality and skills.
• Redesigned homepage: A new homepage showcases what's possible with persona demos and video walkthroughs.
• Fresh visual identity: Updated logo, favicons, and branding across the product.

Two improvements to help you understand and manage running assistants faster.

• Live server logs: A new logs panel on the assistant detail page streams your worker's output in real time, with color-coded log levels and auto-scrolling.
• Playback controls: Start, stop, and clear the log stream without leaving the dashboard.
• Faster config changes: Updating your model or API key now takes effect in about 30 seconds, down from a full redeploy that could take several minutes.
• Visual update status: The dashboard shows an "Updating configuration..." indicator while changes are being applied.

Your assistant is no longer limited to a single messaging platform. You can now deploy to Telegram, Slack, Discord, or any combination of the three.

• Slack support: Connect your assistant to a Slack workspace using a Bot Token and App Token. A built-in setup guide walks you through creating the Slack app.
• Discord support: Connect your assistant to a Discord server with a single bot token. The setup guide covers bot creation, permissions, and server invitation.
• Multi-channel deployments: Enable multiple channels on a single assistant. The detail page shows all active channels at a glance.
• Step-by-step setup guides: Each channel includes an expandable guide with the exact steps, scopes, and permissions needed.

Until now, editing assistant files meant working through openclaw UI or redeploying from scratch. The new file editor gives you direct access to your assistant's files right from the dashboard.

• Browse your assistant's files: Navigate directories, open files, and explore your assistant's configuration and workspace from a familiar file explorer.
• Edit with syntax highlighting: A full-featured code editor with automatic language detection for JSON, TypeScript, JavaScript, YAML, Markdown, shell scripts, and more.
• Save & redeploy in one click: Edit a file and redeploy your assistant in a single action, no need to leave the dashboard.
• Safe by default: Unsaved changes are flagged before closing, and saves are validated to prevent conflicts.

This update improves how managed assistants preserve state, so restarts are less disruptive and recovery is more dependable.

• Better data safety: Backups now run more consistently, including periodic full backups to reduce the chance of drift over time.
• Smoother restarts: Assistants now recover newer saved state more reliably when containers restart.
• More complete recovery: Restores now cover assistant state, workspace files, and installed skills.
• Clear backup health: Admin status now shows last successful backup, last backup attempt, and recent backup errors when present.
• Customer value: Less manual recovery work, fewer surprises after restarts, and more confidence that assistant progress is preserved.

We made this change because cloud browser automation setups were too fragile and required too much manual configuration. The goal of this release is to give you safer defaults, fewer deployment steps, and more reliable assistant behavior in isolated environments.

• Why this matters: Cloud environments can fail in ways that are hard to debug when browser connectivity is not configured securely and consistently.
• What you now get: Deployments now apply secure connectivity defaults automatically, so assistants come up with fewer manual steps.
• Reliability improvement: Isolated sandboxes now default to the Playwright CLI browser path, which avoids common extension-relay failures.
• Stability improvement: Gateway configuration is normalized for current OpenClaw runtimes, reducing startup and connectivity issues.
• Customer value: Faster time-to-working assistant, less deployment troubleshooting, and safer browser automation defaults out of the box.

We are introducing first-class support for fully customizing your openclaw.json, alongside making assistant configuration easier to understand and more dependable.

• Customize openclaw.json: You now have direct control to customize your openclaw.json configuration just how you want it.
• Clearer config status: You can now quickly see which settings are currently applied versus still pending.
• Easier troubleshooting: Side-by-side config comparisons make it faster to spot what changed.
• More reliable updates: Config saves are now applied more consistently across assistant redeploys.
• Cleaner comparisons: Runtime-managed values are filtered out so you can focus on user-controlled changes.
• Supported OpenClaw version: 2026.2.19.

This release focused on product clarity and assistant configuration flow.

• Model updates are easier: Change your assistant model directly from the assistant details page.
• Cleaner logged-in experience: Returning users are now guided to their assistants right away.
• Improved onboarding content: New comparison content helps users choose the right path faster.

This update focused on clarity and usability.

• Smoother account flows: Improved key flows so setup and activation feel more reliable.
• Better docs navigation: Updated structure and links so it is easier to find what you need.
• Clearer guides: Expanded and refined help content for common usage patterns.

This release made external integrations more complete while improving day-to-day stability.

• More complete integrations: Added broader support for assistant management tasks.
• Easier setup: Updated guides and behavior for a smoother integration experience.
• Stability improvements: Fixed issues that could affect account, billing, and assistant operations.

This patch focused on making managed usage more dependable.

• More accurate credits: Resolved edge cases around balance updates and top-ups.
• More stable assistant performance: Fixed issues that could interrupt managed assistant responses.

A full pass on the landing experience improved readability and conversion-oriented messaging.

• New page structure: Improved layout and section flow to make the value proposition clearer.
• Visual and copy polish: Updated visuals and messaging to improve readability and conversion.

This was a major milestone for managed assistant billing and reliability.

• Managed credits are live: Users can now run assistants with platform-managed credits.
• Improved account isolation: Credits are now separated per user for better consistency.
• Updated pricing structure: Managed assistant plans were clarified and separated from other offerings.
• Deployment reliability improvements: Managed assistant deployment flow is now more dependable.

Payments and production reliability moved from setup to functional rollout.

• Payments are live: Added paid plans and credit purchase flows for users.
• Sign-in fix: Resolved a Google sign-in issue that affected some production users.
• Reliability improvements: Fixed several production issues to improve overall stability.
• Legacy branding cleanup: Removed lingering legacy naming in UI/docs and standardized branding.

This release focused on improving security for critical user actions.

• Stronger protections: Added additional safeguards around sensitive platform operations.
• Safer event handling: Improved how background events are validated and processed.
• Improved request checks: Tightened validation for account and assistant-related actions.

This is the first public foundation of the product.

• Core product launch: Released assistants dashboard, authentication, and essential user flows.
• Assistant lifecycle support: Enabled creating, managing, and operating assistants end to end.
• Foundational docs and tools: Shipped initial docs and setup tooling.