Multidim MCP server

multidim-mcp

<!-- mcp-name: io.github.Cherridsaid/multidim-mcp -->

Structured thinking grids for AI agents — a standalone MCP server, pure standard library.

Multidim routes a subject to a set of analysis lenses (a context) and returns a hierarchical grid — axes, sub-lenses, mandatory questions — for the calling LLM to fill in. The thinking stays with the caller: the server provides structure, never cognition. It calls no LLM, makes no network requests, and the same input always produces the same frame.

a filled analysis is checked section by section with actionable error codes.

• Zero dependencies — Python 3.9+, standard library only.
• Deterministic v2 contract — every frame carries a self-verifiable frame_hash;

into every future frame whose subject matches.

• Learned traps — lessons you record once become mandatory questions injected

migrations, backed-up resets, and a guard that refuses to ever touch a foreign ~/.multidim store.

• Hardened store — atomic writes, native cross-process locking, additive

See the difference

An agent analyses "Should we migrate the billing service from MySQL to PostgreSQL?". Every section is filled, every sentence reads fine. Here is what multidim_validate returns on that first pass:

overall verdict: REJECT

REJECT  alternatives        NOT_ENOUGH_ALTERNATIVES, ALTERNATIVE_DUPLICATES_PRIMARY
REJECT  hypotheses          HYPOTHESIS_NOT_FALSIFIABLE
REJECT  second_order_risks  SECOND_ORDER_REPEATS_FIRST
REJECT  cross_talk          GENERIC_DENSITY_HIGH
REJECT  synthesis           SYNTHESIS_WITHOUT_REFERENCES
WARNING premortem           PREMORTEM_SIMILAR_TO_RISKS

The only alternative restated the hypothesis, the hypothesis carried no test that could prove it wrong, the second-order effect repeated the first one word for word, and the conclusion referenced none of the work above. None of that is visible when you read the answer; all of it is reported here, by name.

Redo the rejected sections and the same checker returns ACCEPT. Edit the frame to delete the rule you find inconvenient, and it refuses the whole submission — the frame carries a hash of its own content.

Full transcript, including what the fixed sections look like and what this deliberately does not check: DEMO.md. Reproduce it in one command: python demo.py.

Quickstart

pip install multidim-mcp        # from PyPI
pip install .                   # or from a source checkout

Register the server with any MCP client (stdio transport):

{
  "mcpServers": {
    "multidim": {
      "command": "multidim-mcp"
    }
  }
}

Or run it directly: python -m multidim_mcp, or without installing: uvx multidim-mcp.

The server is listed in the official MCP Registry as io.github.Cherridsaid/multidim-mcp.

Tools

How it works

Shortened. The full README is on GitHub.