What is MCP Solver?
MCP Solver, by Stefan Szeider, turns an MCP host such as Claude Desktop or Claude Code into a solver programmer. The host picks a backend, gets modelling instructions and a persistent IPython kernel with the solver library loaded, writes a program, runs it against the real solver, checks the answer and submits the final program. You keep both the solution and the code that produced it.
Version 4 supports six backends: PySAT, MaxSAT, Z3, CPMpy, Clingo and DIDP. It is MIT-licensed, written in Python, has about 187 GitHub stars, and its last commit landed on 13 October 2026. The earlier v3 design is documented in a SAT 2025 paper and kept on its own branch.
How to install MCP Solver
git clone https://github.com/szeider/mcp-solver.git && cd mcp-solver && uv venv && uv pip install -e ".[agent]"Is MCP Solver worth it?
MCP Solver fixes the most common way a language model fails at combinatorial problems: it guesses an answer that breaks a constraint. Here the model never answers directly. It writes a solver program for one of six backends, PySAT, MaxSAT, Z3, CPMpy, Clingo or DIDP for routing and scheduling, runs it, verifies the result, and only then calls submit_code. The output is a program you can rerun and audit, which matters more than the answer for anyone who has to defend a timetable or a configuration.
The evidence is better than most MCP servers offer. With Claude Opus as host, the README reports 229 of 229 instances solved correctly across CP-Bench (101) and ASP-Bench (128) under strict validation, and all 30 bundled test problems solve with GPT-5.6 Terra through the CLI. These are the author's own runs on benchmarks he built, so read them as strong signals rather than independent results. Version 4 is also a rewrite: the v3 server with fine-grained model-editing tools and MiniZinc is gone, replaced by a kernel and six tools.
The catch: installation and safety both need care. v4 is not on PyPI because a third party holds the mcp-solver name with stale v2.0.0 code, so uvx mcp-solver installs the wrong thing; you must clone and install from source until a name-transfer request is settled. The INSTALL guide says v4 has been tested on macOS only. And the kernel underneath, from the author's agentic-python-coder, executes model-written code automatically in the working directory; its own security notice says to use a VM or container for untrusted inputs.
Our take: for a planner, researcher or engineer with a real scheduling, allocation or verification problem, this is the difference between a plausible answer and one a solver has proved. It is worth the clone and the container. The packaging will catch up; the method is already sound.
Where it earns its place
An operations manager at a 40-bed care home
Build a staff rota that meets shift rules, rest periods and preferences
Describe the rules to Claude; it picks CPMpy or MaxSAT, encodes hard rules and weighted preferences, runs the solver and submits the program. Next month, rerun the saved program with new availability instead of re-prompting.
A firmware engineer
Prove that a bit-twiddling routine matches its specification for all inputs
The agent encodes both versions in Z3 and asks for a counterexample. An unsat result is the proof; a model is a failing input. Run inside a container, since the kernel executes the generated code.
A university course coordinator
Timetable 30 courses into rooms without clashes
Use the Clingo or CPMpy backend to encode rooms, sizes and instructor conflicts; the submitted program becomes next term's timetable generator. Free apart from the host model's tokens.
Before you commit
- Setup
- Clone the repository, uv venv, uv pip install -e ".[agent]", then point your client at uv run --project <path> --extra agent mcp-solver-serve. Do not use the uvx config until the PyPI name transfer completes. Requires Python 3.13 or 3.14 and uv.
- Integrations
- Backends: python-sat, RC2 MaxSAT, z3-solver, cpmpy (with ortools), clingo and didppy. Tools: select_backend, python_exec, python_reset, python_status, python_interrupt and submit_code, plus guide and template resources. The CLI and benchmark harness run through OpenRouter.
- Data and privacy
- Runs locally over stdio. Problem descriptions and generated code go to your host's model provider. Solver programs run in a local IPython kernel in the working directory without sandboxing; optional statistics are written to a JSONL file you choose.
- Cost at scale
- Free under MIT. The cost is the host model's tokens; CLI runs bill OpenRouter usage to your key.
- Limits
- Each CLI run stops after 30 agent steps by default. No Python 3.15 support yet, because ortools and python-sat publish no wheels for it. python_interrupt stops a runaway solve.
What’s good
- Six solver backends from SAT to dynamic programming
- Leaves a verified, rerunnable solver program
- 229 of 229 benchmark instances solved in the author's runs
- No API key; the host's model does the solving
What’s not
- Not installable from PyPI; the name holds stale v2.0.0 code
- v4 tested on macOS only so far
- Kernel runs model-written code with no sandbox
- Needs Python 3.13 or 3.14 and uv
Best for
Operations researchers, engineers and analysts who want an agent to solve scheduling, allocation, configuration or verification problems with a real solver and keep the program.
MCP Solver alternatives
MCP Solver: common questions
- Is MCP Solver worth it?
- The right idea for scheduling and puzzle problems a model should never solve in its head, held back for now by a checkout-only install and a kernel that runs whatever code the model writes. We rate it 3.9 out of 5.
- How much does MCP Solver cost?
- Free and open source (MIT). The server needs no API key; the optional CLI uses your own OpenRouter key.
- What are the downsides of MCP Solver?
- Not installable from PyPI; the name holds stale v2.0.0 code; v4 tested on macOS only so far; kernel runs model-written code with no sandbox; needs Python 3.13 or 3.14 and uv.
- Who is MCP Solver best for?
- Operations researchers, engineers and analysts who want an agent to solve scheduling, allocation, configuration or verification problems with a real solver and keep the program.
- How much does MCP Solver cost for a team or in production?
- Free under MIT. The cost is the host model's tokens; CLI runs bill OpenRouter usage to your key.
- Is MCP Solver safe for company data?
- Runs locally over stdio. Problem descriptions and generated code go to your host's model provider. Solver programs run in a local IPython kernel in the working directory without sandboxing; optional statistics are written to a JSONL file you choose.
- What are the best MCP Solver alternatives?
- The closest alternatives we have reviewed are Lean LSP MCP and Sequential Thinking (reference server).
Answers checked Nov 7, 2026.