mcp-bash-server

The problem

Cowork (and Claude Desktop) run in a sandbox and can't execute commands on your own machine. This server bridges that gap — it connects Cowork to your local bash so Claude can run shell commands directly on your computer, then read their output.

Related MCP server: Bash MCP Server

What it is

A tiny Model Context Protocol server that lets an MCP-aware client run shell commands on your local machine. It exposes one tool, run_bash, and communicates over stdio.

From the client's side, running a command is just a tool call. On this side, the command runs via bash -c with your user's permissions, and its exit code, stdout, and stderr are returned.

⚠️ Security

By default this is an open shell: whatever connects can run anything you can. Only connect clients you trust, and consider using the allowlist below.

Install (easy way — double-click, no terminal)

Download mcp-bash-server.mcpb and double-click it. Claude Desktop opens an install dialog — review the permissions, optionally set an allowlist and default working directory, and click Install. That's it: no Terminal, no pip, no config files. Claude Desktop manages the runtime for you, and because the server is registered locally, it also becomes available inside Cowork sessions.

An .mcpb (MCP Bundle) is the one-click install format for local MCP servers, like a browser extension. This bundle uses the uv runtime, so you don't even need Python pre-installed.

The rest of this README is the manual route, for development or if you'd rather run it yourself.

Requirements (manual route)

• Python 3.10+

• bash on your PATH

Install

cd mcp-bash-server
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Run

The server supports two transports:

python server.py                      # stdio (default) — for the Claude Desktop bridge
python server.py --transport http     # HTTP — serves MCP at http://127.0.0.1:8001/mcp

• stdio is the default. You normally don't run it by hand — the client launches it for you (see "Connect it to Claude Desktop" below). This is the simplest path: Claude Desktop runs the server as a subprocess and bridges it into Cowork automatically.

• http serves the MCP streamable-http transport natively (no supergateway needed). Use it when you want to register the server as a custom connector by URL. Change host/port with --host / --port.

Connect it to Claude Desktop

Add an entry to your claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "bash": {
      "command": "/absolute/path/to/mcp-bash-server/.venv/bin/python",
      "args": ["/absolute/path/to/mcp-bash-server/server.py"]
    }
  }
}

Use absolute paths. Restart Claude Desktop; the run_bash tool should appear.

Connect it as a custom connector (HTTP / URL)

If you'd rather register it by URL (e.g. for Cowork's custom connector flow), run the HTTP transport and point the connector at it:

python server.py --transport http --port 8001
# -> serves http://127.0.0.1:8001/mcp

Then add a custom connector with URL http://localhost:8001/mcp. To keep it running across reboots, manage it with a process manager (e.g. pm2, launchd, or systemd).

The tool

run_bash(command: str, cwd?: str, timeout?: int) → text with exit code, stdout, stderr.

• command — the shell command, run via bash -c.

• cwd — working directory (defaults to where the server started).

• timeout — seconds before the command is killed (capped by BASH_MCP_MAX_TIMEOUT).

Configuration (environment variables)

Example — restrict to a safe set:

{
  "mcpServers": {
    "bash": {
      "command": "/absolute/path/to/.venv/bin/python",
      "args": ["/absolute/path/to/server.py"],
      "env": { "BASH_MCP_ALLOWED": "git,ls,cat,grep,npm,node,python3" }
    }
  }
}

Building the .mcpb yourself

npx @anthropic-ai/mcpb pack . mcp-bash-server.mcpb

This validates manifest.json and produces the installable bundle. The uv server type means dependencies are resolved from pyproject.toml at install time — nothing is vendored, so the same bundle works on macOS, Windows, and Linux.

License

MIT — see LICENSE.