OpenClaw adds an experimental adviser that can flag an agent’s wrong turns

OpenClaw’s v2026.10.10 release gives users a new way to spot an agent going off course while a task is still running. An experimental Advisor can suggest corrections, alongside a redesigned task-details view. The release notes also introduce voice, remote-folder and connected-app options, each with setup requirements or limits that shape how it can be used.

Watching a task before it finishes

The new Chat Details panel collects session information and task progress, replacing progress and pull-request bars above the message box. On wide chat panes, it opens beside the conversation. An automatically opened panel hides as the pane narrows, though users can still open it manually over a narrow conversation.

Users can also inspect running and finished subagents—the workers handling delegated parts of a task—and read their transcripts without editing them. If a run uses tools but ends without a final reply, its saved progress remains visible with an Interrupted label.

Advisor adds a different kind of oversight. OpenClaw says it can review ongoing work and suggest a brief correction when it detects a clear wrong turn, such as repeating an unsuccessful approach or drifting from the request. Users enable it in Settings → System → Labs and can adjust its review intervals. Those checks use extra model calls; they can miss problems or suggest unnecessary changes.

Speaking to the chat agent, working on another computer

The additional browser Talk option separates voice processing from the model answering the request. Split mode uses the same AI model as typed chat, while separately configured services transcribe speech and generate spoken output. Spoken requests go to the current chat agent; read-aloud and dictation also use the configured speech services.

Both speech output and transcription must be configured first, and the browser needs microphone permission in a secure context. Split mode has no realtime voice switching or camera input. Leaving it unset preserves existing realtime Talk. Authenticated Talk follows ordinary tool permissions and approvals rather than requiring a second spoken confirmation for actions already permitted.

Agents can now work directly in a folder on another connected computer. OpenClaw could run on a server while working with documents or a project on a laptop, without copying the folder to the server. These conversations do not support attachments or the Files panel. The paired computer’s settings can restrict selection to permitted folder roots or disable device-folder sessions entirely.

Connecting apps without opening the whole service

The optional MCP Relay connects ChatGPT or Claude to OpenClaw conversations without an internet-facing Gateway port or sharing the Gateway password. The plugin is off by default. Pairing uses Gateway sign-in or a single-use code, and users choose and revoke each app’s access in Connected apps.

That connection has a privacy boundary: the relay can read content passing through it. OpenClaw advises granting access only to trusted apps and a trusted relay. Full role access can include administration and approvals.

Finally, supported events from a connected MCP server can start an automation instead of waiting for a timer to check for changes. The connected app need not stay open. This opt-in, experimental integration requires a public HTTPS callback, with only the callback route exposed—not the rest of OpenClaw.

Support is limited to the OpenAI webhook profile, not every draft MCP Events feature. Receiving an event does not mean the agent completed its task, and an event lost upstream may not be recoverable.