ResiMind MCP
MCP server for exact algebra: checks polynomial identities, expands expressions and keeps rules that passed verification
- 292 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Runs a local Python process and parses your expressions with its own parser instead of executing them as code
- The learn_polynomial tool changes rule memory inside the process; the other four only read
Install
- Works in
Manual install
git clone https://github.com/1105216375-alt/resimind.git && cd resimind && git checkout v0.16.0 && python3.11 -m venv .venv && .venv/bin/python -m pip install '.[mcp]'MCP needs Python 3.11 or newer. Installation may download build tools.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Runs a local Python process and parses your expressions with its own parser instead of executing them as code
- The learn_polynomial tool changes rule memory inside the process; the other four only read
This is third-party code. Review the repository files before installing.
What it does
ResiMind gives an agent five local tools for polynomial work. They run over MCP stdio, and the same calls are available in a console mode that reads JSON lines. No model key is needed and computation runs offline, so the network is used only for installation. verify_identity compares the coefficients of two expressions and, when they differ, returns the exact mismatch the agent can use to repair its answer. expand_polynomial expands a long expression into sparse terms with exact numerators and denominators as decimal strings. learn_polynomial solves an example and tries to extract a rule from it, while recall_polynomial applies an already admitted rule without a model. list_rules shows the accumulated rules and where they came from. The repository also ships a skill, verify-and-repair-algebra. It teaches the agent to check derivation steps in order, name the first rejected step and try at most three repairs. A verified result means the coefficients matched exactly; it is not a Lean proof.
Who it is for. For people who ask Claude Code, Cursor or Claude Desktop to derive formulas and expand brackets and want the result backed by an exact coefficient check.
Good fit when
- An agent expanded (x+1)**2 into x**2+1 and you need to catch it: verify_identity returns the coefficient mismatch
- A long product of polynomials has to be expanded exactly, with no rounding: expand_polynomial returns terms with numerator and denominator as strings
- A multi-step derivation is checked step by step and you need the first wrong one
- The same expansions recur: learn_polynomial may store a session rule, which you check through learning.new_rule_ids
- You are checking an integration and want to see the server reject a false identity, accept its correction and apply a learned rule
Not a fit when
- You need calculus, functions or a variable in a denominator: the server accepts only integer literals, parentheses, addition, subtraction, multiplication, nonnegative integer powers and division by a nonzero integer constant
- You need a formal theorem proof: exact coefficient checking and a Lean proof are different kinds of evidence, and the server does not run Lean
- Rule memory has to survive a restart: in MCP mode it lives only in the process, and saving is available from Python through the lifelong adapter
- The expression exceeds the limits: 8,192 characters, 16 variables, degree 32 or 4,096 terms
- You need a production-proven product: the author marks the project as experimental
Example request
Expand (2*x+3*y)**3 with ResiMind and check that my expansion 8*x**3+36*x**2*y+54*x*y**2+27*y**3 is correctMCP
- Transport
- stdio
- Authentication
- not required
Limitations
The server handles rational-coefficient polynomials within defined size limits. Functions, variable denominators and calculus are unsupported. `unresolved` means an unsupported operation, a limit reached or no matching rule; it does not disprove an identity.
Rules live in process memory and are lost on restart. Matching coefficients do not imply that a new rule was learned: check `learning.new_rule_ids`. Persistent storage is available through a separate Python interface.
MCP needs Python 3.11 or later. Version 0.16.0 is experimental and uses MIT. The agent sees expressions and answers, so local calculations do not make the whole conversation offline. Model integrations and Lean, planning or support demos are separate from the five MCP tools.
How to disable. Remove the resimind block from mcpServers in the client configuration and restart the client. The server does not edit client settings or start an HTTP service, so you can then delete the clone folder with its .venv virtual environment.
FAQ
How do I connect ResiMind to Claude Code?
Install the package from a clone with python -m pip install '.[mcp]' in a Python 3.11 environment and add a server whose command is the .venv Python with the arguments -m resimind mcp. Launching it by hand shows nothing because a stdio server waits for protocol messages.
What do verified, rejected and unresolved mean?
Verified means the equality or expansion passed an exact coefficient check. Rejected means the expressions have different coefficients and the claim is false. Unresolved means the operation is unsupported or limits ran out, which is not a disproof.
Does it need a model key or internet access?
No key is needed and the tools run offline. The network is used only during installation. The model_calls counter in each reply shows that no model was called.
How do I turn on rule learning?
Call learn_polynomial on two compatible examples, for instance (a+b)**3 and (u+v)**3. If the schemas align, the learning.new_rule_ids field gets a rule id and recall_polynomial applies it to (2*x+3*y)**3.
Can learned rules persist between sessions?
Not in MCP mode: memory lives in the process and is cleared on restart. Saving and loading exist in the Python interface of the lifelong adapter, and the core rechecks rules when they are reloaded.
Discussion
No questions or reviews yet.
Related
MCP serversAn MCP server and CLI where deterministic tools verify every agent claim about code and binaries against ground truth
A local code graph with version history: an MCP server that tells the agent who calls a function and what an edit breaks
One MCP server gives agents a code map across over 300 languages, git history, shared memory and document search with OCR.
Roslyn CodeLens MCP gives an agent a semantic map of a .NET solution: callers, type hierarchies, DI, diagnostics and safe renames
An MCP server that builds a code graph of functions, classes, calls and routes and answers the agent's structural queries locally
The CodeAlive MCP server: the agent searches large codebases by meaning and exactly and follows call relations without a full read
README in short
Release v0.16.0 includes source code, a wheel and an MCPB bundle. Outside MCP, `python -m resimind api` in the installed environment provides a JSON Lines interface. It is a separate protocol and cannot be configured as a JSON-RPC server.