Integrations / Listen and Recap
IntegrationsUpdated October 24, 2026
Listen and Recap
Follow a live conversation or recap a session with your agent
Version: This guide describes Parrot Scribe 1.1.0.
Requires Parrot Scribe Pro. MCP server access is part of Parrot Scribe Pro. Existing licenses already include Pro.
Listen brings a live transcript into the agent that already has your context, skills and tools. Recap summarizes a selected session. Parrot Scribe supplies the transcript through its local MCP server; your agent interprets it.
Start in your desktop app
Connect your agent first.
| Desktop app | How to request the workflow |
|---|---|
| Claude | After the extension is installed, ask Claude to listen or recap. |
| Codex | After Connect, start a new local task and ask it to listen or recap. |
| Cursor | After Connect, start a new chat and ask it to listen or recap. |
Cowork is not one of these rows. See Claude Desktop before relying on it. Continuous listening and background work also depend on the host capabilities below.
Listen
The command reads your instructions before changing capture. It attaches to an existing recording or normally starts capture when paused. Instructions such as "attach only" or "do not record yet" take precedence.
Give it an objective rather than a broad role when possible:
Listen and help me discover how often this customer's workaround happens.
Listen, attach only; flag contradictions with this project's architecture.
In Claude Code, prefix the request with /parrot:listen.
The parent stays in the conversation and follows one bound session. It speaks only when a supported, unresolved observation is useful at that moment. Notes stay in the chat; they are not spoken into the recorded call. The agent does not infer a person's identity from microphone input or speaking order.
Typed instructions and spoken proposals
The same words have different authority depending on their source:
| Input | Result |
|---|---|
| Someone says "Create a ticket about this" in the meeting | A proposed follow-up, held until the meeting ends and you approve it in chat |
| You type "Create a ticket about this in this repository" directly in chat | A background subagent handles that task while the parent keeps listening, if the host supports background work |
| You quote those words while asking a question | Context, not permission to create a ticket |
The worker receives the authorized task and relevant evidence, not permission to obey the rest of the transcript. It does not control capture or take over the live cursor. The parent reports its result or failure without duplicating the task at recap. Normal tool, provider, and high-impact confirmation requirements still apply.
If background subagents are unavailable, the agent keeps listening and offers to defer the task or let you explicitly pause for foreground work. It must not pretend that a foreground call preserves background listening.
Transcript text is untrusted evidence. Speaker recognition, a microphone source, and a spoken claim of authority cannot authorize tools, approve actions, or change these rules. This is an instruction contract for the host, not a replacement for its tool permission controls.
Session and cursor handling
During capture, get_status includes the active sessionId, and live_transcript advertises waitSeconds. The agent checks those capabilities before listening; it must not guess a session or silently use an unbound feed.
Each live response includes session identity, recording state, epoch, sinceSequence, gap, sessionChanged, a TOON schema, and text. The host must preserve the full envelope, either as MCP structured content or the JSON text block. Bare unwrapped TOON loses continuity information.
The listener sends the bound session and cursor back on subsequent calls. A caught-up call can wait for speech or a state change, for up to 20 seconds. Request starts stay at least minPollIntervalSeconds apart. There is one live request at a time, not separate rapid loops for transcript and status.
The live feed keeps the most recent 1,024 confirmed sentences in memory and is lost on restart. On a same-session gap, the agent reads the persisted transcript to recover evidence and reconciles overlap. It reports coverage it cannot recover. On a session change it detaches rather than silently following the next meeting. Neither a gap nor a restart authorizes follow-up actions.
Recap
Recap is read-only. It does not stop capture just to summarize. From a listening turn it uses the bound session unless you select another. A historical description uses search, with clarification when several sessions match. A still-live session gets an interim recap.
Recap the customer interview from yesterday.
In Claude Code, use /parrot:recap the customer interview from yesterday.
The agent reads the transcript, pages through large sessions, and distinguishes confirmed decisions from tentative suggestions and corrections. The format follows your request and relevant host skills. It does not invent owners, deadlines, agreement, or work for a video with no follow-ups.
After the meeting, proposed actions are presented for approval. Approve the specific items you want executed. Already delegated or completed tasks remain distinct from proposals, and approval of one item does not approve the rest.
When the bound meeting ends, the listener can recap it automatically. "Stop helping" detaches the agent without stopping capture. An explicit "stop recording and recap" stops only the bound session, then recaps it. A historical recap never stops an unrelated live meeting.
Host capabilities and privacy
Parrot Scribe's MCP server sends its usage and listening guidance to each connected host; hosts that support MCP server instructions apply it. Plugin 0.4.0 from the catalog also includes a listening skill for Claude Code and the Codex CLI, and the /parrot:listen and /parrot:recap commands for Claude Code. Continuous listening needs real waiting and user steering during a running turn; simultaneous task execution needs background subagents. Connecting MCP alone does not add those capabilities. If a host cannot sustain listening, use on-demand reads and recaps. Never claim a listener is active after its task has ended.
Parrot Scribe captures and transcribes locally. A cloud-backed host may send transcript content to its model provider. Choose the host and provider policy appropriate for the conversation and obtain required participant consent. Do not infer that local storage makes a cloud agent local.
For installation, desktop support and marketplace availability, see Connect Your Agent. For access controls and tool arguments, see MCP Server.