fip-mcp

An MCP server that turns community FAIR Implementation Profiles (FIPs) into guidance and deterministic compliance checks for AI agents such as Claude Code.

A FIP records a community's agreed FAIR choices: identifiers, repositories, formats, vocabularies, licences and provenance models. It is published as nanopublications by the FIP Wizard. Today FIPs are mostly filled in once and rarely read again. This server makes them operational. An agent helping a researcher asks the FIP what the community agreed before it suggests anything, and can check work against the FIP.

fip-mcp was developed in the OSCARS-funded FIESTA project (FAIR Image Data Workflows in Galaxy), and is maintained by the Science Live community.

Tools

Design rules:

• Verdicts are code, not LLM judgement. Each declared resource maps to a probe of facts read from the target. A resource with no probe is manual, so the agent asks the user instead of guessing.
• A community gap stays a gap. no-choice is reported as a gap in the FIP, and the agent is told not to put its own preference in its place.
• Latest version. The FIP Wizard republishes a FIP and its declaration index, and both supersede their predecessors. The resolver only follows successors that are themselves FIPs.

Install and use

pipx install fip-mcp
claude mcp add fip -s user -- fip-mcp     # as an MCP server
fip-check check fiesta-eo 10.5281/zenodo.20960690   # same operations from the shell
fip-check crosswalk fiesta-bio fiesta-eo

Nanopubs are cached in ~/.cache/fip-mcp. They are immutable, so they are cached for good. Version lookups always go to the network.

For development: pip install -e '.[dev]' and python -m pytest -q (the tests need no network).

Status

Prototype. The short names in src/fip_mcp/data cover the Astronomy, Bioimaging and Earth Observation FIPs of the OSCARS-FIESTA community FAIR Image Data Workflows in Galaxy. Any tool also accepts the nanopublication URI of another FIP, but the check probes have so far only been exercised on FIESTA outputs.

Assistant test

eval/ tests whether an AI assistant follows a community's FIP, with and without fip-mcp. It asks 6 questions (repository, data format, licence, identifier, vocabulary, provenance) for each of the 3 FIESTA communities, 3 runs per condition. The expected answers are read from the published FIPs, not typed by hand. Results from 2026-10-20 (Claude Opus 5.5, web and file tools disabled) are in eval/results/:

python eval/run_eval.py --runs 3 --out eval/results/<name>     # needs Claude Code
python eval/score.py eval/results/<name>

Citation

Please cite fip-mcp with the metadata in CITATION.cff.

Funding

fip-mcp was developed within FIESTA, a project funded by OSCARS (Open Science Clusters' Action for Research and Society). OSCARS is funded by the European Union's Horizon Europe programme under grant agreement No 101129751. Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them.

Science Live is a community-driven, unfunded initiative. Further funders will be listed here as they support the work.