Show HN: Constraint Programming in Your AI Agent

An MCP server that exposes MiniZinc constraint solving and optimization to LLM clients such as opencode, Claude Desktop, and Cursor. It lets an agent parse, type-check, and solve MiniZinc models directly from a chat session.

Built with the MCP Python SDK v2 and the MiniZinc Python binding.

Only two things need to be installed, once per machine:

• uv —curl -LsSf https://astral.sh/uv/install.sh | sh
• MiniZinc 2.6+ with theminizinc executable onPATH (includes a default solver, Gecode)

Everything else is fetched automatically by uv — there is no clone, no venv setup, and no manual pip install on your side.

Install it globally (best if you use it in several projects):

uv tool install --from git+https://github.com/carban/minizinc-mcp minizinc-mcp

Or run it on demand each time, with nothing installed:

uvx --from git+https://github.com/carban/minizinc-mcp minizinc-mcp

The server runs over stdio. Tell your MCP client to launch it:

opencode — project level (add this to opencode.jsonc in your project):

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "minizinc": {
      "type": "local",
      "command": ["uvx", "--from", "git+https://github.com/carban/minizinc-mcp", "minizinc-mcp"]
    }
  }
}

opencode — global (add the same mcp.minizinc block to ~/.config/opencode/opencode.json):

{
  "mcp": {
    "minizinc": {
      "type": "local",
      "command": ["uvx", "--from", "git+https://github.com/carban/minizinc-mcp", "minizinc-mcp"]
    }
  }
}

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "minizinc": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/carban/minizinc-mcp", "minizinc-mcp"]
    }
  }
}

Restart your client. Six tools should now be available, prefixed with minizinc_:

• minizinc_list_solvers
• minizinc_validate_model
• minizinc_solve_model
• minizinc_solve_model_by_path
• minizinc_get_model_info
• minizinc_get_flatzinc

Quick sanity check — ask your client: "list the available MiniZinc solvers". You should see gecode, chuffed, highs, and anything else installed on the machine.

The result is a JSON object like:

{
  "status": "OPTIMAL_SOLUTION",
  "objective": 9,
  "solution": { "objective": 9, "x": 9, "y": 1 },
  "statistics": { "time": 0.204, "nodes": 3, ... }
}

status is one of SATISFIED, OPTIMAL_SOLUTION, ALL_SOLUTIONS, UNSATISFIABLE, UNKNOWN, or ERROR. validate_model and solve_model never raise in normal operation — errors are returned inside the result dict.

Clone the repo, then:

uv sync          # create the environment and install mcp + minizinc

The server speaks the MCP stdio transport, so it is launched as a subprocess by an MCP client. Run it with the SDK inspector:

uv run mcp dev server.py

that opens the MCP Inspector in the browser where every tool can be called interactively. A minimal programmatic smoke test:

uv run python -c "
import asyncio
from mcp import Client
from mcp.client.stdio import StdioServerParameters

async def main():
    params = StdioServerParameters(command='uv', args=['run', 'python', 'server.py'], cwd='.')
    async with Client(params) as client:
        result = await client.call_tool('solve_model', {
            'model_code': 'var 1..10: x; var 1..10: y; constraint x + y = 10; solve maximize x;'
        })
        print(result.content[0].text)

asyncio.run(main())
"

Install the test dependencies, then run the suite:

uv sync --group dev
uv run pytest -q

The tests in tests/ launch the server end-to-end over stdio and call every tool through the MCP protocol, solving the example model in example/. They need a working MiniZinc install (the same prerequisite as for developers).

• params follows JSON representation: JSON arrays map to MiniZinc arrays; numbers, strings, and booleans map to their native MiniZinc types. Exotic types like sets and enums are not fully expressible this way.
• Do not combine all_solutions withmax_solutions ; the MiniZinc driver rejects the combination.
• MiniZinc requires a solver that supports the model (e.g. chuffed /gecode for CP,highs /cbc for MIP models). Uselist_solvers to see what is installed.
• Solutions are returned inline in the tool result; read_only_hint is set on all tools, so they do not modify your files or system.