Connect AI assistants (MCP)

What MCP is

The Model Context Protocol is an open standard for connecting AI assistants to tools and data. An MCP server offers tools - named actions with a description and typed arguments - and an MCP client (the assistant) decides when to call them while it works on your request. BenchMesh is an MCP server: its tools read the bench, handle flows and recordings, and set instruments within the access you grant.

Endpoint

http://127.0.0.1:57666/mcp

• Streamable HTTP with JSON responses (POST JSON-RPC; no server stream, so GET /mcp answers 405). Protocol versions 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05.
• The app must be running - the endpoint is part of it. There is nothing to install.
• The desktop app listens on port 57666, or on a free port if that one is taken. The endpoint actually in use, and the bridge command for this computer, are shown in Automation → Settings → MCP. For the stdio bridge, set BENCHMESH_URL (e.g. http://127.0.0.1:57700) when it is not on 57666.

Access levels

You choose what clients may do in Automation → Settings → MCP access:

A client that calls a control tool at read access gets a message that the user can allow it in those settings.

Tools

Instrument tools take device (the id from bench_overview), class and channel (default 1). Status values come back as numbers in base units (V, A, W).

Setup

Claude Code

claude mcp add --transport http benchmesh http://127.0.0.1:57666/mcp

Claude Desktop

Claude Desktop starts its servers as programs, so it uses BenchMesh’s stdio bridge: the BenchMesh executable started with --mcp opens no window and forwards the messages to the running app. Add this to Claude Desktop’s configuration (Settings → Developer → Edit Config, the file claude_desktop_config.json) and restart Claude Desktop:

{
  "mcpServers": {
    "benchmesh": {
      "command": "<path to the BenchMesh executable>",
      "args": ["--mcp"]
    }
  }
}

Windows installs BenchMesh for your user; the path is usually the one below - the exact path for your computer is shown in Automation → Settings → MCP. Note the doubled backslashes in JSON.

{
  "mcpServers": {
    "benchmesh": {
      "command": "C:\\Users\\<you>\\AppData\\Local\\BenchMesh\\benchmesh.exe",
      "args": ["--mcp"]
    }
  }
}

Linux: the path of the AppImage file itself, wherever you keep it (it must be executable).

{
  "mcpServers": {
    "benchmesh": {
      "command": "/home/<you>/Applications/BenchMesh_<version>_amd64.AppImage",
      "args": ["--mcp"]
    }
  }
}

If BenchMesh is not running, the bridge answers every request with an error asking you to start the app.

Cursor and other HTTP clients

Clients that speak Streamable HTTP connect to the URL directly - for Cursor in ~/.cursor/mcp.json (or the project’s .cursor/mcp.json):

{
  "mcpServers": {
    "benchmesh": {
      "url": "http://127.0.0.1:57666/mcp"
    }
  }
}

Headless server

With benchmesh-server instead of the desktop app, the same bridge is benchmesh-server mcp; BENCHMESH_URL points it at the server (default http://127.0.0.1:57666):

{
  "mcpServers": {
    "benchmesh": {
      "command": "benchmesh-server",
      "args": ["mcp"],
      "env": {"BENCHMESH_URL": "http://127.0.0.1:57666"}
    }
  }
}

Or try a tool by hand - any HTTP client without a browser Origin header works:

curl -s http://127.0.0.1:57666/mcp -H 'content-type: application/json' \
  -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/call",
       "params": {"name": "read_instrument", "arguments": {"device": "psu-1", "class": "PSU", "field": "VOUT"}}}'

Safety

• Read by default. A connected client can look, not touch, until you choose control.
• Energising needs explicit agreement. Even at control, a setter that can switch an output or input on needs confirm_energize: true and a flow that does needs arm: true. The server tells clients to ask you before passing either; without it the call is refused.
• Confirmed like the UI. set_instrument waits until the instrument reports the new value, and flows run with all of the flow safety rules - cleanup, safe state, bounded loops, one run at a time.
• Stopping is always allowed: abort_run works at read access, and you can abort from the app at any time.
• This computer only. The MCP endpoint is served on 127.0.0.1 only, by the desktop app and by benchmesh-server alike. Remote access (a browser on your local network, HTTPS with a password) never reaches it: AI clients get control only on the bench computer. The cross-site guard refuses requests from web pages of other sites - a page you visit cannot reach the endpoint through your browser, as the MCP specification requires.

Example conversations