// the take
Jacobian is an MCP server (plus CLI and Python API) that exposes a library of small, typed mathematical operations — exact linear algebra, number theory, graph algorithms, SAT/SMT via Z3, polynomial algebra — instead of one big do-everything math tool. It's aimed at people building AI agents that need to discover and call a specific mathematical operation (via math.find/math.run) rather than prompt a model to "just solve" something.
The find-then-run pattern (inspect a typed contract, then call it) is a sane answer to the usual MCP problem of either dumping a huge tool schema on the model or hand-rolling a thin wrapper around SymPy yourself. It's built on real backends (SymPy, Z3, Python-FLINT, NetworkX) rather than reimplementing math, and it's explicit about exactness vs. approximation instead of quietly returning floats where you expected a rational. The project also has a genuinely serious benchmark/verifier setup (conjecture-probe datasets, per-task verifiers, CI across adapters) rather than a handful of toy examples, which is unusual for a repo this size.
It's 0.23.0 and says plainly that operation contracts can still change between releases — not something to wire into a production agent pipeline yet. The tested install contract is narrow: CPython 3.12/3.13 on glibc Linux x86-64 only, with Alpine/musl explicitly called out as broken for the full backend stack, so Docker/Alpine users or Windows-native devs are outside the supported lane. Singular and QEPCAD, needed for some operations, are optional system runtimes installed outside the normal pip/npm path, which is an easy thing to not notice until an operation fails. At 311 stars with a scope this broad (SAT/SMT, algebra, geometry, topology), it's early — expect sharp edges and a smaller pool of people who've already hit your specific bug.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →