Using discopt from Claude Desktop (MCP server)#
discopt mcp runs a Model Context Protocol
server on stdio. Any MCP client can then build and solve discopt models on your
machine: Claude Desktop, Claude Code, or another one.
Install and register#
pip install 'discopt[mcp]'
discopt mcp --print-config
The second command prints the entry to merge into Claude Desktop’s
claude_desktop_config.json. On macOS that file is
~/Library/Application Support/Claude/claude_desktop_config.json; on Windows it
is %APPDATA%\Claude\claude_desktop_config.json. The entry looks like this:
{
"mcpServers": {
"discopt": {
"command": "/path/to/your/python",
"args": ["-m", "discopt.cli", "mcp"]
}
}
}
It names the absolute path of the Python that has discopt installed. Desktop
does not inherit your shell’s PATH, so a bare discopt command is often not
found there. Restart Desktop after editing the file.
For Claude Code, register the same command:
claude mcp add discopt -- "$(python -c 'import sys; print(sys.executable)')" -m discopt.cli mcp
Tools#
Tool |
What it does |
|---|---|
|
Runs Python source that builds a |
|
Solves a |
|
Describes a model’s size and structure without solving it |
|
Converts between |
|
Reports the version, the install location and the Python running the server |
A solve returns the same fields as discopt solve --format json: status,
objective, bound, gap, gap_certified, wall_time, node_count and the
solution x. x is truncated to max_vars entries (default 200), and when that
happens x_truncated says how many entries there were. time_limit defaults
to 60 s per call. A result is a proven global optimum only when status is
optimal and gap_certified is true.
How it runs#
Each request runs in a fresh worker process, never in the server process. There are three reasons:
Clean protocol stream. The stdio transport is the server’s stdout, so anything a solver or a model script prints there would corrupt it. The worker redirects its own stdout to stderr and returns its result on a private channel.
Hard timeout. A worker still running 60 s past its
time_limitis killed, and the call returns an error instead of hanging the client.Crash isolation. A crash takes down the worker, not the server.
Errors are returned as {"ok": false, "error": ..., "stderr_tail": ...}, with
the tail of the worker’s traceback.
Warning
solve_python and .py files execute arbitrary Python as the user running
the server. That is how a client writes a model, and it is no more than a
client with a shell tool could do. Register the server only with clients you
would trust with a shell.