Workbench MCP Server

An MCP server for remote access to a project over the network. It lets a Claude CLI running on another machine read/write files, search code, run allow-listed commands and work with git in the context of a chosen project — and also launch a Claude CLI on the host machine in the background (claude_start).

Originally built for Workbench, but it works with any project — the path is set via MCP_PROJECT_ROOT.

Quick start

git clone https://github.com/jazz-max/workbench-mcp.git && cd workbench-mcp
npm install
cp .env.example .env
# edit .env: set MCP_AUTH_TOKEN and MCP_PROJECT_ROOT
npm start

On a successful start you'll see:

MCP server listening on http://0.0.0.0:3100/mcp (auth enabled)
Project root: /path/to/your/project

Install via npm / npx

Without cloning the repository:

# one-off run, no install
MCP_AUTH_TOKEN=<token> MCP_PROJECT_ROOT=/path/to/project npx workbench-mcp

# or install globally
npm install -g workbench-mcp
MCP_AUTH_TOKEN=<token> MCP_PROJECT_ROOT=/path/to/project workbench-mcp

When run via npx/global install there is no .env next to the server — set the configuration via environment variables directly. In this case MCP_PROJECT_ROOT is required (otherwise the server doesn't know which project to serve).

Run via pm2 (recommended for long-running use)

pm2 start server.mjs --name workbench-mcp
pm2 save
pm2 startup          # start on boot (one-time)

pm2 logs workbench-mcp     # logs
pm2 restart workbench-mcp  # restart after changes

Configuration (.env)

Variable	Default	Description
MCP_PORT	3100	Server port
MCP_AUTH_TOKEN	(empty)	Bearer token. Without a token the server runs without authentication (use only on a trusted local network)
MCP_PROJECT_ROOT	parent directory	Absolute path to the project the server serves

Generate a token: node -e "console.log(crypto.randomUUID())"

Connecting Claude CLI (from another machine)

# host IP (macOS): ipconfig getifaddr en0
claude mcp add workbench \
  --transport http \
  --url http://<IP>:3100/mcp \
  --header "Authorization: Bearer <TOKEN_FROM_.env>"

claude mcp list      # verify
claude mcp remove workbench

Tools

Tool	Description
project_info	Project overview: CLAUDE.md, servlet list (if servlets.json exists), docs/, current branch
read_file	Read a file (path relative to project root), up to 1 MB, with line numbers
write_file	Write a file (creates directories). Forbidden: .env*, node_modules/, vendor/
list_files	List a directory (optionally recursive, name filter), up to 500 entries
search_code	Search by content (regex/grep), glob filter, excludes node_modules/vendor/.git/storage/public
run_command	A command from the allowlist, execFile (no shell), 30s timeout, output up to 1 MB
git_status	Current branch, git status --porcelain, last 10 commits
claude_start	Launch claude -p in the background with a prompt → returns a taskId
claude_result	Status and result of a background task by taskId (running/done/error)

allowlist for run_command: git, php, composer, npm, node, npx, ls, cat, head, tail, find, grep, rg, wc, diff, file, which, pwd, echo

Scenario: remote development from two machines

Machine A — the project and the MCP server. Machine B — the Claude CLI for exploration.

                    LAN (port 3100)
  Machine B  ──────────────────────>  Machine A
  Claude CLI                          MCP server
  (research, specs)                   (project files, git, commands)

Note: Claude Code CLI must be installed on both machines — on B as the client that connects to the MCP server and drives the work, and on A, where claude_start runs Claude locally on the host for coding.

A typical flow for a new feature:

1. The MCP server is running on machine A.
2. In the Claude CLI on machine B:

   > Use project_info to understand the project structure
   > Read app/Servlets/BaseServlet.php — the base class
   > Read app/Servlets/DemoBooksScraper.php — an example
   > Research example.com and write a spec to docs/spec.md
   

3. Without switching to machine A, ask your Claude to delegate the coding to the host:

   > Ask claude on the workbench to write the parser per docs/spec.md
   

   Claude on machine B calls claude_start → a separate Claude CLI launches on machine A ("the workbench") and writes the code. Collect the result via claude_result.

Security

• Requests are protected by a Bearer token (MCP_AUTH_TOKEN). Without a token the server runs without authentication — only expose it on a trusted network.
• run_command is restricted to an allowlist and uses execFile (no shell injection).
• write_file forbids writing to .env*, node_modules/, vendor/.
• The server grants remote access to files and command execution — do not expose it to the public internet without a separate reverse proxy with TLS and authentication.

Troubleshooting

• Cannot connect: both machines on the same network; check the IP; curl http://localhost:3100/mcp should return 405; check the firewall (macOS may ask to allow Node.js).
• Unauthorized: the token in --header must match MCP_AUTH_TOKEN.
• Command not allowed: the command is not in the allowlist.
• Port in use: lsof -ti:3100 | xargs kill -9.

License

MIT.