Skip to main content
Messaging now keeps more of a conversation intact across the places people already talk to their Claw. Telegram gains richer messages and media, Slack keeps live progress and the final answer together, Discord adds opt-in Activities and voice rooms that understand who is present, and the native apps keep media and pending sends inside the conversation where they belong. Across supported channels, OpenClaw now holds accepted messages through managed restarts, reports whether a connection is usable, recovering, or blocked, and preserves an uncertain send instead of blindly sending it again. Recovery starts once OpenClaw has accepted the message, and each service still controls what it can confirm beyond that point.
On supported channels, messages OpenClaw has accepted now stay pending through a managed restart, and channel status shows whether a connection is usable, recovering, or blocked. When a send times out without a confirmed result, OpenClaw keeps that outcome uncertain and can warn on the next contact rather than creating a likely duplicate. Recovery begins after local acceptance.Eligible single-choice questions can use native controls on Telegram, Discord, and Slack, while longer Telegram and Discord turns can show a short status headline with compact tool activity. Multi-select and free-text questions continue through the supported client or text path, and Telegram partial-answer streaming remains a separate opt-in mode.
On rich-enabled Telegram accounts, agents can send native details, tables, checklists, math, maps, file references, locations, venues, and a compatible round video note. Large nested replies paginate within Telegram’s limits, and rejected rich structures fall back to the complete plain-text answer instead of dropping the useful part.Busy conversations hold together better too. Follow-ups stay ordered during active work, accepted updates can resume after a restart, abandoned delivery claims stop freezing the chat, and verified public poll activity returns to the chat or forum topic where it began. Anonymous polls remain display-only, and quoted text reaches the agent as attributed quotation rather than active instructions.
Discord now supports opt-in Activities that open configured OpenClaw widgets directly inside Discord. Voice agents can also see who is in the room, use different wake-name rules for one-on-one and group conversations, and optionally join only while a human is present. Activities require explicit channel configuration, a Discord client secret, and a public HTTPS route; occupied-room auto-join is a separate option and does not change the existing always-on default.Message retries now reuse their identity to reduce duplicates, interaction replies settle in order, stale reply context is discarded, and repeated resume failures can recover without restarting all of OpenClaw. Forum-thread sends still remain uncertain when retrying them could produce a duplicate.
Slack progress now stays in one conversation from the first status through the final answer, keeping the running commentary, work steps, and answer inside one streamed message unless an operator chooses the more compact alternative. Charts and tables can render natively too, with readable text preserved when Slack cannot use the native form.One organization-installed app can also serve the Enterprise Grid workspaces Slack grants it, with messages, actions, approvals, and proactive delivery bound to the right workspace. Destinations outside the current conversation require an explicit workspace, requests without verified workspace identity are rejected, and relay mode plus org-wide surfaces remain unsupported.
WhatsApp can now list groups from the linked account without making someone hunt through logs or invite links, and it avoids opening a competing connection while OpenClaw already owns that account. If OpenClaw restarts after accepting an incoming message, the pending work can continue without later events jumping ahead, while multipart replies keep the parts and receipts that actually succeeded instead of replaying the entire response after one part fails.
Signal replies now keep their native quote block through ordinary, chunked, media, and durable delivery when quoting is enabled. Messages received just before a crash can resume from local storage, and a failed recipient is reported instead of disappearing into a false success. Group sends remain conservative after any member receives the message, so one failed recipient does not cause the whole reply to be repeated to everyone else.
iMessage is now an official installable plugin that carries existing configuration and state through the move out of core. Eligible approval requests can use native Messages polls, while older bridges, SMS, and failed poll sends keep the existing text, reaction, or command fallback.Remote Mac setups also keep attachments with the correct existing chat and can use supported remote paths and actions without sharing a filesystem with the machine running OpenClaw, while incoming messages already stored before a crash can replay. New setups still need the plugin and imsg on a signed-in Mac, and this remains separate from the iOS and macOS OpenClaw chat apps.
In Feishu, an agent can resend a sticker previously received by that bot account and, when an operator adds labels, find the right sticker by keyword. Accepted Feishu messages and comments can also resume after a restart, streaming cards keep the latest accepted version, and unsupported image formats remain available as file attachments instead of being discarded. Sticker actions and labels are opt-in and do not retroactively organize an existing collection; the sticker and comment work is specifically verified for Feishu rather than every Lark path.
Mattermost can now read bounded, paginated history from channels the bot can already access once an operator enables the feature. Integrations that rewrite or cancel replies now complete that work before delivery, preventing an early preview from exposing content that the hook removes or changes. Completed sends also keep the real Mattermost post identity across previews, actions, media, and durable paths.
Matrix replies can now use native spoilers, underline, tables, and up to 100 discoverable room or personal custom emotes. More importantly, an encrypted room now fails visibly when encryption is unavailable instead of quietly falling back to plaintext text, attachments, filenames, or media metadata. Large tables can still choose their own readable formatting fallback without weakening the room’s confidentiality boundary.
Configured Microsoft Teams approvers can approve or deny eligible exec and plugin requests from a native Adaptive Card in the chat or channel thread where the request began, then see the recorded outcome on that same card. Quoted replies, attachments, streaming results, and duplicate handling also stay bound to the originating conversation. The cards cover configured exec and plugin requests only and require valid authentication, permissions, tenant authorization, and routing.
Google Chat webhook events accepted by OpenClaw now remain queued through restarts, keep their order within each space, and recognize Google’s retries before they create duplicate work. Recovery begins when the Gateway admits the event; events Google never delivered remain outside that recovery path.
LINE keeps mixed rich replies, their quick actions, and media together, using the reply path for as many as five messages when the response fits instead of consuming Push API quota unnecessarily. Long code that will not fit a Flex card falls back to ordinary chunks, unsupported media remains visible as a link, and blank or rejected rich structures no longer take the whole reply with them.In groups that require a mention, slash commands addressed to the bot now run as commands, and quoting one of its recent Gateway-sent messages can address it without a second mention. Date, time, and rich-menu selections reach the agent, while supported actions and choices return as native controls with visible text for anything LINE cannot render.A bot can also post one room-specific introduction when it actually joins an allowed group, while respecting both channel-wide and per-account opt-outs. Webhook events are stored before acknowledgement so work OpenClaw already admitted can continue after a restart. Affected installations that passed through the brief pre-drain queue transition migrate eligible accepted rows before delivery resumes, while genuinely mismatched rows remain quarantined and messages for removed accounts wait until that account is restored.
The Twilio-backed channel can now send and receive MMS and show recent provider or carrier states such as sent, delivered, failed, or conflicted without retaining message bodies or phone-number addresses. SMS and RCS messages OpenClaw has already accepted can survive a restart in sender order, while replay protection fails closed instead of accepting work it can no longer protect from duplication.Delivery observations are kept for 30 days and reflect the latest state reported by Twilio or a carrier, which can differ from recipient-visible delivery. If the replay cache fills, new events are rejected rather than accepted without duplicate protection, and Twilio does not retry that response by default.
ClickClack can place a team discussion beside an OpenClaw session, giving people somewhere to coordinate around the work without turning the agent’s main transcript into a meeting room. Guided and command-line setup, readable discussion names, native command menus, attachments, optional group mention rules, and opt-in progress make that room easier to use while keeping the final answer visible.Opening a discussion still requires an authorized operator and a reachable ClickClack deployment. Generated names remain best effort, and mention-gated or bot-to-bot conversations have to be enabled deliberately.
Trusted Claws can talk directly through Reef, OpenClaw’s bundled end-to-end encrypted agent channel, using friend-code pairing, terminal registration and friendship controls, and discovery that keeps an external peer distinct from a local thread. Operators can now add plain-language inbound and outbound sharing rules for their own sensitive topics and named friends, while deterministic denials for secrets and credentials remain in force and changing the rules invalidates pending approvals from the old policy.When an inbound message needs owner review, that recorded decision now owns later redelivery instead of rerunning the guard until the answer changes, and later inbox messages can keep moving while it waits. Temporary inbound guard failures leave the item parked for another attempt rather than rejecting it, with approved delivery receiving one final guard check and outbound sends still failing fast.When a peer rejects a message, the sending agent gets bounded feedback and one controlled chance to rephrase before it stops for owner guidance instead of arguing in a loop.
OpenClaw can now expose selected agents to explicitly trusted external agent systems through the A2A v1.0 protocol. Configured peers can discover those agents, submit and poll authenticated tasks, receive replies as artifacts, and exchange text or structured data, while an unconfigured plugin registers no discovery or task routes at all.This first version supports one account, keeps tasks in memory, and does not yet provide streaming, push notifications, or cancellation. Authenticated peers currently operate as trusted callers rather than passing through the normal command-policy decision, so this is a deliberate interoperability path for known peers rather than a public agent endpoint.
Buzz is now an official OpenClaw channel for team rooms, with guided setup that can route different rooms to different agents and live directories that give those agents current room and member names without replacing the stable UUIDs used for automation. One Gateway can now run named Buzz accounts with separate bot identities, credentials, rooms, routing, and lifecycles while preserving the legacy root account and environment fallback, and updating one account does not unnecessarily disconnect healthy siblings.Replies can use native mentions and threads, and operators can choose flat automatic replies and typing while explicit message-tool and CLI targets remain explicit. Mention-gated rooms can carry a small amount of recent authorized context into the next turn without running the model on every background message.Setup finishes only after Bot-role membership is verified, a room name must resolve uniquely before it can be used safely, native mentions cap at 50, and passive context remains opt-in and tightly bounded.
QQ Bot has moved to Tencent’s integrity-pinned external plugin while keeping the public qqbot channel ID and carrying existing credentials, account selection, allowlists, approval restrictions, streaming behavior, and group tool policy when Tencent 2.0 can represent them safely. If an older policy cannot be translated without weakening it, Doctor stops for an explicit repair instead of quietly changing the rules. Inbound envelopes already accepted into the local queue can also resume after a restart.
Zalo Bot and Zalo Personal now keep accepted messages through restarts using separate recovery paths. Zalo Bot records a webhook before acknowledging it, while Zalo Personal resumes accepted socket messages from its account queue without one delayed conversation holding up unrelated chats. Polling and webhook modes remain mutually exclusive, and targeting, reactions, formatting, delivery errors, and partial receipts remain specific to the path that handled the message.
Tlon now saves an accepted Urbit message before acknowledging it, so work already admitted to OpenClaw can resume after a restart without immediately abandoning the server cursor. Replies gain native Markdown lists, and oversized SSE events or JSON payloads stop before they can grow in memory without bound.There is one hard recovery limit. If Eyre has definitively deleted a channel, OpenClaw can create and subscribe to another one, but the old cursor and its server-side history are gone.
Nextcloud Talk can now process different rooms concurrently while preserving message order within each room. Up to 32 deliveries can be active at once, with excess work queued, and sends or reactions to an unresponsive self-hosted or remote server now fail within a bounded time instead of hanging indefinitely.
A Nostr direct message can move to the next configured relay when one cannot connect, return the real connection error when every relay fails, and expose the successful relay’s real event ID instead of a synthetic timestamp. Named accounts, protected keys, profile imports, reply targets, and ordered encrypted chunks also stay attached to the selected account more consistently.
Synology Chat now records inbound webhook work before returning success, allowing accepted messages to resume after a restart. If the local write fails, OpenClaw returns an error so the sender can redeliver, while the upstream retry remains controlled by Synology. Long Unicode replies keep their order, lookups and responses are bounded, and uncertain outbound sends remain unresolved instead of being replayed into a duplicate.
Twitch chat already accepted into OpenClaw’s local queue can continue after a process crash. Stalled user lookups can be cancelled, stopped accounts stay stopped even when an earlier connection attempt finishes late, and ordinary replies keep normalized attachment links while internal tool traces and XML scaffolding are removed before they reach chat.
IRC channel messages already admitted to OpenClaw can resume in order after a restart without echoing the bot’s own replies back into the room. Direct-message recovery stays tied to the connection that accepted it and stops if that identity changes, while mentions follow IRC nickname rules, Markdown becomes readable plain text, and internal tool traces stay out of the channel.
Android now writes text, images, and voice notes to its outbox before using the network, keeping them through offline periods and restarts until conversation history confirms delivery. If the outcome remains uncertain, the item stays visible with explicit Retry and Delete controls and is not sent again automatically.Supported audio and video play inline, video uses the native upload flow, and notification replies return to the exact saved conversation or fail. Older 2026.7.x OpenClaw installations use a reduced compatibility path.
The iOS and macOS OpenClaw apps can play supported managed audio and video inside chat, upload video from the native composer, and hand the active attachment to system Now Playing controls. A slow accepted reply stays visibly pending while live state and saved history reconcile, so late history cannot clear a newer turn just because it arrived second.A run that truly produces no output eventually releases the composer without inventing a reply. Native video uploads retain the 20 MB limit, and these chat apps remain separate from the iMessage channel plugin.
Control UI chat now gives supported generated documents and managed audio and video their own named cards instead of exposing raw attachment instructions or treating every file like a download. Ready video can expand over the conversation, compatible delivered media can play inline, and every attempted attachment keeps a visible named outcome with actionable failure guidance. Active or unknown file formats remain unavailable rather than becoming downloadable, while media the browser cannot prepare keeps its download action.Across Control UI, WebChat, and TUI, prompts, live work, attachments, and replies stay aligned when several clients share a conversation or reconnect at different times. Delayed prompts no longer jump below their live replies, and terminal users keep a privacy-safe attachment warning beside the answer and after history reload.These are OpenClaw’s own chat clients rather than external messaging services, and the saved conversation remains the final record when a live update and refreshed history disagree.
Voice Call can opt into an agent’s main transcript when an operator wants phone and desktop conversations to share history, while the default remains one session per phone. Realtime agents can speak their final words and request a hangup, and failed startup, carrier hangup, or media silence now closes the unusable session instead of leaving it active until the maximum duration.Choosing the main transcript places raw call turns in primary history. A carrier can still reject a requested hangup, and carrier, provider-media, Discord, Talk, and Meet calls keep their own credentials and lifecycle rather than being treated as one interchangeable voice system.
Raft setup now says clearly when the machine running OpenClaw cannot find the Raft executable instead of presenting a configured channel as healthy. A passing probe means the command-line tool was available at that moment, not that a later connection or message has already succeeded.