Get Started

Integrating the Pieces MCP with Claude Cowork brings your workflow context directly into Anthropic's autonomous task automation agent. Claude Cowork is a general-purpose task automation agent that runs within Claude Desktop—it reads, writes, and organizes files, executes multi-step tasks, and works across your filesystem.

With Pieces MCP connected, Cowork can use your Long-Term Memory as context while executing tasks—like generating standups from yesterday's work or creating summary documents from your captured workflow data.

**Plans required:** Claude Cowork requires Pro, Max, Team, or Enterprise subscription. **Status:** Research preview (launched January 2026, Windows added February 2026).

Prerequisites

Claude Cowork runs inside Claude Desktop. Install Claude Desktop first. Ensure you have Pro, Max, Team, or Enterprise subscription. PiecesOS must be installed and running. Enable the Long-Term Memory Engine (LTM-2.7) through the Pieces Desktop App or the [PiecesOS Quick Menu](/products/core-dependencies/pieces-os/quick-menu).

Installing PiecesOS & Configuring Permissions

Follow the instructions below for a detailed guide on setting up and configuring PiecesOS to correctly pass captured workflow context to the Pieces MCP server.

Connecting Pieces MCP to Claude Cowork

Cowork uses the same MCP configuration as Claude Desktop. You have two options:

Option 1: Via Connectors UI (Pro/Max/Team/Enterprise, for remote URLs)

For remote access, you need a public HTTPS URL. Set up ngrok or another tunnel first.

Open Claude Desktop. Go to **Settings > Connectors**. Click `Add custom connector` and enter your ngrok URL: ```plaintext https://YOUR_NGROK_URL.ngrok-free.app/model_context_protocol/2025-03-26/mcp ``` Save and restart Claude Desktop.

Option 2: Via JSON Config with stdio Bridge (recommended for local use)

**Recommended when PiecesOS and Claude Desktop are on the same machine.** Uses the localhost URL via an mcp-remote bridge.

Install mcp-remote globally with a pinned version for security:

npm install -g mcp-remote@0.1.38

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "pieces": {
      "command": "mcp-remote",
      "args": [
        "http://localhost:39300/model_context_protocol/2024-11-05/sse"
      ]
    }
  }
}

See the Claude Desktop guide and MCP Bridge for full details.

Using Pieces Tools in Claude Cowork

Once connected, start a Cowork session and Pieces LTM tools are available alongside file system access.

Example tasks:

  • "Use my Pieces Long-Term Memory to find what I worked on yesterday and write a standup to ~/Desktop/standup.md"
  • "Search my Pieces memory for everything I've captured about the authentication redesign and create a summary document in my project folder"
  • "Use my Pieces audio transcriptions from today's meetings and create a meeting notes file with key decisions and action items"

Key Differences from Claude Chat

Feature Claude Chat Claude Cowork
File access No Yes (designated folder)
Task parallelism No Yes (sub-agents)
Autonomous execution No Yes
MCP tools Yes Yes
Session continuity Per-conversation Per-task (no cross-session memory)

Verification

  1. Open Claude Desktop and start a Cowork session (look for the folder/file access UI).
  2. Ask: "What Pieces tools are available?"
  3. Give a multi-step task: "Search my Pieces memory for my recent work on [project name] and write a status update to my Desktop"

Troubleshooting

If you're experiencing issues integrating Pieces MCP with Claude Cowork:

  1. Cowork Option Not Visible: Ensure you have Pro/Max/Team/Enterprise plan.

  2. Pieces Tools Not Appearing: Restart Claude Desktop after config changes.

  3. MCP Tools Unavailable: Confirm the Pieces server is configured in Claude Desktop—Cowork shares Desktop's config.

  4. Windows Not Available: Windows support was added February 10, 2026—update Claude Desktop.


You're now set to enhance your Claude Cowork workflow with powerful context retrieval through Pieces MCP. Happy coding!