Anthropic’s Parka Design Would Turn Meetings Into Claude Agent Tasks
The notable feature is the handoff: spoken follow-ups could become structured work for Claude Cowork, Claude Code, or a person. The product remains unavailable in public Claude Desktop builds.
Story brief
3 key pointsReverse-engineered artifacts from Claude Desktop 1.32885.1 reveal an unreleased Anthropic workflow, codenamed Parka, that could convert recorded meetings into structured assignments routed to Claude Cowork, Claude Code, or a person. The design includes speaker-attributed transcripts, editable notes, owners, full prompts, execution types, and optional Claude session links. It is not a functioning public feature:...
- 01
Parka’s action schema distinguishes cowork, code, and manual follow-ups instead of producing only a meeting recap.
- 02
Meeting records can combine system or microphone audio, calendar metadata, transcription keyterms, live transcripts, summaries, notes, and edits.
- 03
Reverse engineering found a 551-byte native Parka loader stub, while enabling the feature flag produced no screen or meeting-list request.
Anthropic has designed an unreleased Claude Desktop workflow that could turn meeting follow-ups into structured tasks for Claude Cowork, Claude Code, or a person. Application artifacts identify the project by the internal codename Parka and describe a system that moves from recorded conversation to agent-ready assignments.
Parka is not available to users today. Public Claude Desktop builds mark its meetings feature as unavailable, and the packaged native Parka loader is an empty 551-byte stub. In a test that changed the renderer’s feature status to supported, Claude still showed no Parka screen or navigation item and made no observable request for a meeting list. The artifacts therefore show a detailed product design, not a working public meeting tool.
From a recorded conversation to an assignment
The contract starts with meeting capture. Parka can take system audio, microphone audio, or both, and accepts calendar-event metadata and transcription keyterms. It also includes controls to list, retrieve, start, stop, edit, and delete meetings—evidence of a workflow built around individual meeting records rather than a one-off audio upload.
As a meeting proceeds, the design provides for live transcript events attributed to speakers. It also includes summaries, notes, granular editing controls, and error states. That combination places the transcript alongside editable meeting material, with room to correct or organize information before follow-up work is represented as an action.
The action record makes the design different from a summary
Parka’s follow-up schema is more specific than a checklist in meeting notes. Each action can include a title, description, owner, full prompt, execution type, autoRunnable value, and optional Claude session URL. Its three execution types are cowork, code, and manual.
What a Parka follow-up can carry
- An owner, title, and description for the assignment.
- A full prompt and an optional URL for a Claude session.
- An execution type that directs the item to Cowork, Claude Code, or manual work, plus an autoRunnable value.
That structure is the central product signal. Instead of stopping at a recap of who said what, Parka is designed to preserve an assignment’s destination and instructions. An item intended for Claude Code can be represented differently from one meant for Claude Cowork or a human owner, while remaining part of the same meeting-derived record.
A visible blueprint, with the release still unresolved
The Parka details came from reverse engineering Claude Desktop version 1.32885.1 artifacts for macOS and Windows. The contract exposes recording, transcription, editing, and task-routing components, but neither examined public build exposed a usable Parka interface. Whether Anthropic releases Parka, keeps that name, or enables automatic execution remains unresolved.
Editorial analysis
Our Read
Our Read: Parka suggests Anthropic is considering meetings as a source of work for its existing agent products, rather than treating transcription as an endpoint. The consequential design choice is the structured action record: it includes an owner, a full prompt, an execution type, and an autoRunnable field. Those are the ingredients for a handoff, but not proof that the handoff will work safely or accurately. The next meaningful signal would be a public release that shows how people review, edit, assign, and approve extracted tasks before Claude Cowork or Claude Code acts on them.