Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Jacobian

Two MCP tools, math.find and math.run, over a catalogue of more than 1,400 exact maths operations built on SymPy, Z3 and FLINT.

Reviewed by ReviewRamble editors · Updated

Setup docs Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.7/5

An unusually principled maths server that keeps the model's tool list to two entries, though its pace and pre-stable label make it a research instrument more than a dependency.

What is Jacobian?

Jacobian, from the GitHub user morluto, gives an agent a searchable library of typed mathematical operations instead of a code interpreter. The server exposes two tools: math.find matches a described need against the catalogue or returns the exact schema and examples for one operation, and math.run executes one operation with a typed payload and returns a typed result. Coverage runs from exact linear algebra, polynomial algebra and number theory to graphs, bounded SAT and SMT solving, group theory, combinatorics, probability, geometry and topology. The same library is usable from a CLI and as a Python package.

The repository was created in July 2026, has about 370 stars, is MIT-licensed and released 0.23.0 on 3 October 2026 to PyPI and npm, with commits continuing into November. The project calls itself pre-stable: experimental operation contracts may change between releases.

How to install Jacobian

npx jacobian@latest setup

Is Jacobian worth it?

Jacobian's design choice is to give the model mathematical moves, not solvers. In our count of the v0.23.0 source there are more than 1,400 operation declarations across 33 domain packages, each a narrow function such as an extended gcd, an exact determinant, a Galois orbit or a tree-automaton minimisation, with at least one worked example that the test suite executes. Results are exact where they claim to be, and approximate or incomplete answers say so in the result itself. The model decides how to compose them. Hiding the catalogue behind math.find and math.run keeps the context cost to two tool schemas, where listing every operation would mean hundreds of definitions.

What it can touch is close to nothing. According to the source, the local server is stateless and built-in operations neither read nor create server-owned state; in our reading, no tool reads files, makes network calls or runs commands, and it runs over stdio. The practical risk is compute: the docs warn that exact operations may legitimately run for minutes or longer, so a client timeout can cut off a long calculation. The remote mode, jacobian-remote-mcp, binds to 127.0.0.1:8000 by default and refuses to start without a token file, with each bearer token at least 32 characters, unless the development-only --allow-anonymous flag is set. It logs a keyed hash of each math.find query, not the text.

The catch: this is a young, fast-moving, mostly single-author project. One account has made about 3,830 of the commits, pull request numbers passed 4,570 in under four months, and 85 issues were open when we checked. The tested install contract is narrow, CPython 3.12 or 3.13 on glibc Linux x86-64, and Alpine cannot install the full stack. Singular and QEPCAD, needed for a small set of operations, are not provisioned for you.

Our take: for exact answers rather than plausible ones, Jacobian is a sharper instrument than letting a model write SymPy on the fly, and it complements Lean tooling such as lean-lsp-mcp, which checks proofs rather than computing values. For a constraint model you want to write yourself, mcp-solver remains the closer fit.

Where it earns its place

  • A PhD student in algebraic combinatorics

    Check a conjectured identity for Littlewood-Richardson coefficients on many small cases

    The agent finds the coefficient and Kostka-number operations with math.find, runs them across partitions up to a chosen size and reports any counterexample with the exact inputs. Free; the limit is how long each case takes.

  • A verification engineer at a 40-person hardware start-up

    Ask whether a set of bit-vector constraints from a spec is satisfiable

    The agent encodes the constraints for the bounded SMT operation, which calls Z3 directly, and returns a model or an explicit unknown. It saves writing a one-off Z3 script, though production checks should still live in your own test suite.

  • A maths teacher preparing a linear algebra problem set

    Produce exact answers and determinants for 30 rational matrices

    The agent runs matrix.determinant.compute and related exact operations on each matrix and formats the results, with no floating-point rounding in the answer key.

Before you commit

Setup
With Node.js 20.17+ and uv installed, npx jacobian@latest setup detects your agents and writes their MCP entries and a jacobian-math skill after you review the changes (--dry-run previews). Or run uvx --python 3.12 --from jacobian jacobian-mcp, or pip install jacobian.
Integrations
Two MCP tools (math.find, math.run), an operation://catalog resource, a jacobian CLI with inspect and run, and a native Python API. Backends are SymPy, NetworkX, Z3 and Python-FLINT, with optional Singular and QEPCAD.
Data and privacy
Local stdio server with no stored state and no file or network access in its tools; your problems and results go to your model provider. Remote mode requires a token file (32+ character tokens) or an explicit anonymous flag, defaults to 127.0.0.1, and logs only an HMAC hash of discovery queries.
Cost at scale
Free under MIT. Costs are model tokens and CPU time; the docs say exact operations may run for minutes or longer.
Limits
Pre-stable 0.23.0; experimental contracts change between releases. Python 3.12 to 3.13 only (PyPI requires under 3.14); musl and Alpine unsupported. Long computations can outlast client timeouts. 85 open issues.

What’s good

  • Only two tool schemas in context
  • More than 1,400 exact, typed operations
  • No file, network or shell access in the tool surface
  • Remote mode fails closed without bearer tokens

What’s not

  • Pre-stable; contracts may change between releases
  • Mostly one author, 85 open issues
  • Tested only on Python 3.12 and 3.13, glibc Linux x86-64
  • Some operations need Singular or QEPCAD installed separately

Best for

Mathematicians, researchers and engineers who want an agent to compute exact algebraic, combinatorial or logical results rather than approximate them in prose.

Jacobian alternatives

Jacobian: common questions

Is Jacobian worth it?
An unusually principled maths server that keeps the model's tool list to two entries, though its pace and pre-stable label make it a research instrument more than a dependency. We rate it 3.7 out of 5.
How much does Jacobian cost?
Free and open source (MIT). Self-hosting a remote server is at your own infrastructure cost.
What are the downsides of Jacobian?
Pre-stable; contracts may change between releases; mostly one author, 85 open issues; tested only on Python 3.12 and 3.13, glibc Linux x86-64; some operations need Singular or QEPCAD installed separately.
Who is Jacobian best for?
Mathematicians, researchers and engineers who want an agent to compute exact algebraic, combinatorial or logical results rather than approximate them in prose.
How much does Jacobian cost for a team or in production?
Free under MIT. Costs are model tokens and CPU time; the docs say exact operations may run for minutes or longer.
Is Jacobian safe for company data?
Local stdio server with no stored state and no file or network access in its tools; your problems and results go to your model provider. Remote mode requires a token file (32+ character tokens) or an explicit anonymous flag, defaults to 127.0.0.1, and logs only an HMAC hash of discovery queries.
What are the best Jacobian alternatives?
The closest alternatives we have reviewed are MCP Solver, Lean LSP MCP and Sequential Thinking (reference server).

Answers checked Nov 16, 2026.