Capsule Bash MCP

An MCP server that gives your AI agent the ability to run bash commands in a secure, persistent, sandboxed environment.

How It Works

Each session runs inside a WebAssembly sandbox. The sandbox provides:

• Persistent state: cwd, env vars, and filesystem changes persist across commands within a session

• Filesystem diff: every run response includes a diff of what changed on disk

• Isolated memory: each session gets its own address space, no cross-session leakage

• No host access: the sandbox cannot reach your host filesystem or network

Learn more about Capsule Bash.

Tools

Sessions

Commands within the same session_id share cwd, environment variables, and filesystem state across calls.

Example

Ask your AI agent:

"Write a Python script that averages a list of numbers."

The agent calls run sequentially:

{ "command": "mkdir -p /data && cd /data", "session_id": "custom_session" }
{ "command": "echo 'nums = [x for x in [1, 2, 3, []] if isinstance(x, int)]\nprint(sum(nums) / len(nums))' > avg.py", "session_id": "custom_session" }
{ "command": "python3 avg.py", "session_id": "custom_session" }

Each call returns stdout, stderr, exitCode, a filesystem diff, and the updated state to enrich context and keep trace in conversation history.

Setup

Add to your MCP client configuration (e.g. Claude Desktop, Cursor):

{
  "mcpServers": {
    "bash": {
      "command": "npx",
      "args": ["-y", "@capsule-run/bash-mcp"]
    }
  }
}

Limitations

• Not all bash commands and options are implemented. Feel free to open an issue if a command is missing or behaves unexpectedly.