claude-bridge-mcp

Reach the Claude Code CLI on one machine from MCP clients on another.

claude-bridge-mcp is a small MCP server that runs next to an authenticated claude CLI and exposes it over HTTP with server-sent events (SSE). A remote MCP client connects to /sse, calls a tool, and the bridge spawns claude locally and returns the JSON result. It is for a developer who wants their own agents or MCP clients on other machines (a laptop, a VPS, a home server) to reach the Claude Code install on their main machine over a private network.

The bridge drives the official CLI under your own login. What a Claude plan allows is set by Anthropic's terms, so check them before you give other people or services access to a bridge.

What it does

• claude_execute runs a prompt through claude --print --output-format json in an allowed working directory. Default timeout 120 s, overridable per call.

• claude_query is the same call with a 60 s limit and --verbose, meant for questions. Read-only is not enforced by the bridge; what the CLI may do is decided by your local Claude Code permission settings.

• claude_read_file reads a text file inside the directory allowlist. Paths go through realpath first, so a symlink cannot point outside the allowlist.

• claude_git_status returns the branch, git status --porcelain and the last five commits of a repository inside the allowlist.

• Access control: optional bearer token (constant-time compare), IP allowlist (localhost always passes), directory allowlist, and at most 5 SSE connections per IP.

• Load control: FIFO queue with configurable concurrency (default 2), queue depth capped at 50, queue timeout, a 100,000-character prompt limit, and output capture that stops at about 1 MB per stream.

• Observability: one log line per request and tool call, GET /health (no auth) and GET /metrics.

Related MCP server: claudecode-mcp

How it's built

MCP client --GET /sse, POST /messages--> Express 4 + MCP SDK (SSE transport)
                                              |
                                              +-- tools.ts --> executor.ts --spawn--> claude CLI (local login)

TypeScript (ES modules), Express 4, @modelcontextprotocol/sdk and zod for config. Four source files, about 630 lines. There is no test suite in the repo.

Repository layout

• src/index.ts: HTTP server, auth and IP middleware, SSE and message routes, health and metrics.

• src/tools.ts: the four tool definitions and handlers.

• src/executor.ts: CLI spawn, queue, timeouts, working-directory check.

• src/config.ts: environment parsing and defaults.

• smithery.json: marketplace listing.

• .env.example: configuration template.

Run it locally

Needs Node 18+ and the claude CLI on PATH, already logged in.

git clone https://github.com/0motionguy/claude-bridge-mcp.git
cd claude-bridge-mcp
npm install
npm run build && npm start        # or `npm run dev` for watch mode
curl http://localhost:3100/health

Point an MCP client at it with npx -y mcp-remote http://<host>:3100/sse.

The server reads plain environment variables and does not load .env itself. Export them, or on Node 20.6+ start it with node --env-file=.env dist/index.js.

The defaults are open: the server binds all interfaces and the token is optional. Before exposing it beyond localhost, set BRIDGE_API_TOKEN and BRIDGE_ALLOWED_IPS and bind to a private interface. The directory allowlist covers the working directory and the two file tools; it does not sandbox the spawned CLI.

Status

Beta. A single-purpose tool at version 1.1.0 with five commits, last push 2026-02-17. It is not released from this repository to npm: the claude-bridge-mcp name on the npm registry belongs to an unrelated project, so install from a clone.

Licence: MIT, see LICENSE.