What it does
Multidim routes a subject to a set of analysis lenses, called a context, and returns a hierarchical grid of axes, sub-lenses and mandatory questions. The calling model does the reasoning; the server provides only structure. It makes no LLM calls and no network requests, and the same input always produces the same frame. Each frame carries a hash, so a filled analysis can be checked section by section, and rejected sections are redone within a set number of rounds. Lessons recorded as learned traps become mandatory questions in future frames whose subject matches.
The validator checks for falsification tests on hypotheses, alternatives that differ from the primary one, second-order effects distinct from first-order ones, a pre-mortem that does not copy the risk list, a synthesis that references real identifiers, and filler-phrase density. The four seed contexts are generic, code_review, technical_writing and decision, and a context can be created or extended with keywords, axes and traps.
The server exposes four tools: multidim_analyze, multidim_contexts, multidim_validate and multidim_learn. It runs over stdio with any MCP client, directly with Python or through uvx. Its store is created on first run in a per-user data path, which an environment variable can override. Writes are atomic and locked across processes, and a corrupt store is backed up before any reset.
Limitations and requirements
Requires Python 3.9 or newer; the project uses only the standard library. The server does not reason itself, so a calling LLM is needed to fill the frame. Validation checks structure and rules, not the correctness of the content. A guard refuses any read or write that would resolve into a foreign ~/.multidim store.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
Multidim
Structured thinking grids for AI agents: deterministic analysis frames, validation, learned traps
v0.2.0As shown on registry.modelcontextprotocol.io when we last read it