MathKernel: an MCP server that hands agents math they can trust
162 typed tools, six trust levels, and Lean-verified proofs — a fresh MIT release that offloads the one thing LLMs reliably botch: exact math.
Copy markdown
The LLM plans, the kernel proves
MathKernel flips the usual split: the model parses intent while a typed backend does the computation and carries the evidence. It exposes 162+ tools (all prefixed `math_`) over MCP across 50+ domains, from calculus to Kalman filters. Point your agent at it and arithmetic stops being a guess.
Every answer ships with a trust label
Results carry one of six levels — formal (Lean 4 + Mathlib), exact, symbolic (SymPy), interval-certified (mpmath), high-precision, or plain numeric — and overall trust is capped by the weakest step. A lone decimal like 0.1 downgrades the whole result, so nothing gets silently promoted to 'proven.'
Five engines behind one interface
It routes to SymPy, Z3 and Lean 4, mpmath intervals, custom exact-integer code, and NumPy/SciPy/CUDA, keeping a full provenance DAG that survives a restart. Coverage spans graph algorithms, finite fields, control systems, SDEs, and PDEs with certificates.
Wire it into Claude Code or Cursor this weekend
It's MIT-licensed and on PyPI: `pip install 'mathkernel[mcp]'`, then run `mathkernel-mcp` (Lean + Mathlib auto-install on first launch). Fair warning: it's a brand-new single-developer v1.3.0 — test it before you trust the 'formal' badge in production.