The era of manual meeting notes might be over. A newly discovered leak inside the Claude Desktop macOS app reveals an unreleased feature internally codenamed Project Parka. Rather than simply transcribing your meetings, Parka is designed to listen to your calls and directly assign the resulting action items to AI agents like Claude Code.
Here is a breakdown of what we know about Anthropic's ambitious new workflow.
How Project Parka Works Under the Hood
Reverse-engineered macOS packages reveal that Parka is a Mac-first feature currently hidden behind a production kill switch. It is built to handle the entire meeting lifecycle:
- Audio Capture: Takes in system audio, microphone audio, and calendar-event metadata.
- Live Transcription: Streams speaker-attributed transcripts in real-time while generating summaries and editable notes.
- Task Routing: Converts spoken follow-ups into structured assignments for Claude Cowork, Claude Code, or a human.
Instead of merely summarizing what was said, Parka translates the conversation into actionable, runnable work.
The Action Schema: Where It Gets Real
The secret sauce lies in Parka’s highly structured follow-up schema. Each extracted task contains specific instructions that act as a handoff to an AI agent.
Based on the leaked design, here is a conceptual example of what the action schema looks like under the hood:
{
"title": "Update main tech stack",
"description": "Refactor the backend services to use the new authentication flow.",
"owner": "Claude Code",
"fullPrompt": "Implement the updated JWT auth strategy in auth.ts according to the meeting discussion.",
"executionType": "code",
"autoRunnable": true,
"sessionUrl": "claude://session/..."
}
The inclusion of an autoRunnable field is particularly striking. It suggests that tasks generated during a noisy meeting could immediately spin up AI workflows without requiring human intervention.
The Competitive Threat
This leak positions Anthropic in direct competition with AI meeting assistants like Granola, Notion AI, and Otter. While most tools stop at recapping who said what, Parka treats the meeting as an automated prompt generator for its existing ecosystem of agents.
Currently, Parka remains an empty native stub. However, if Anthropic brings this design to reality, it will fundamentally change how developer teams handle post-meeting workflows.
How do you feel about an AI agent automatically writing code based on what was casually discussed in a meeting?
