Feature request: fully featured Cowork MCP server — Claude can drive NotebookLM via MCP but not Anthropic's own Cowork

Feature request

Ship an official, fully featured Cowork MCP server so that Claude on any surface — Claude Code, claude.ai, other Cowork sessions — can manage a user's Cowork estate on their behalf: create and drive sessions, read and search their history, and retrieve their outputs.

The absurdity, concretely

Through community MCP servers I can sit in Claude Code and have Claude fully operate Google's NotebookLM for me — create notebooks, add sources, run queries, generate and download artifacts. A third-party product, driven end-to-end by Claude over MCP.

Meanwhile Claude cannot see or touch Anthropic's own Cowork, running on the same machine, on the same account. Asking Claude Code "what happened in my latest Cowork session?" requires reverse-engineering an undocumented session store — and for cloud sessions there is nothing to reverse-engineer, because no transcript exists on the machine at all. Anthropic's ecosystem interoperates with everyone's products except Anthropic's.

Scope: what "fully featured" means

Read (the acute pain):

• list_sessions / read_session / search_sessions across local and cloud sessions
• Access to session outputs and artifacts

Write/drive (the actual prize):

• Create a Cowork session in a given project/space with selected folders
• Send a task/message to a session and read the response
• Resume or continue an existing session
• Manage scheduled tasks / Dispatch programmatically

This would let users compose their own cross-surface workflows: Claude Code kicking off a Cowork quote-preparation task overnight, a Cowork session asking Claude Code to run git operations, a supervisor session reviewing all venture sessions weekly. MCP is Anthropic's own protocol — Cowork is the most conspicuous product in the world not speaking it.

Why this is nearly free to build

The infrastructure already exists. The desktop app registers the user's machine as a remote device, proxies locally-registered MCP servers into cloud Cowork sessions (mcp__remote-devices__*), and syncs workspace folders bidirectionally. Session transcripts are already stored and served to the app from the backend. Scheduled/cloud execution already exists. An official cowork MCP server is a thin wrapper over surfaces Anthropic already operates.

Current state, for the record

1. Local Cowork sessions store full transcripts on disk (~/Library/Application Support/Claude/local-agent-mode-sessions/...), but the format is undocumented and the directory layout has changed between app versions, silently breaking any tooling built on it.
2. Cloud Cowork sessions leave no transcript on the machine at all. The only local traces are a pointer file (remote-session-spaces.json) and the files the session writes into its workspace folders. The conversation is viewable only inside the desktop app window. No documented export or API path (the claude.ai data export's coverage of Cowork cloud sessions is undocumented).
3. Recent app versions default new Cowork tasks to the cloud, and the composer in my current build (v1.25927.0) offers no visible local/cloud chooser — so the no-local-transcript case is becoming the default case.

To keep an auditable record of my own sessions I have had to: parse the undocumented local store, diff remote-session-spaces.json from a launchd watcher to detect cloud sessions starting, and add instructions to every workspace's context files ordering sessions to hand-transcribe their own conversations into the synced folder. That works, but it is fragile, best-effort, and clearly not how this should be done.

Ask

• An official fully featured Cowork MCP server (bundled with the desktop app or installable), covering read and write/drive scopes above; or as interim steps
• A read-only session MCP or documented export path that includes cloud sessions, and
• A stability commitment (or changelog) for the local session store layout.