MCP & scientific tools (policy)

Castalia Institute — Mag.AI-Science

This program treats Model Context Protocol (MCP) as the default pattern for connecting AI workflows to real scientific systems: instruments, literature and citation graphs, lab information systems, data catalogs, and analysis pipelines—never as unconstrained root access.

Principles¶

1. Explicit scope — Each integration declares what it can read or act on.

2. Least privilege — Credentials and tokens are minimized and rotated per cohort policy.

3. Auditability — Tool calls are logged where policy requires; students document scopes in artifacts.

4. Human gates — High-stakes actions (instrument runs, external submissions) require explicit human confirmation unless faculty publishes an exception for a sandbox.

Typical surfaces (illustrative)¶

Examples of what MCP-backed tools might connect to—not an exhaustive or guaranteed list:

• Literature and metadata APIs (e.g. search, retrieval, structured bibliographic data)

• Instrument interfaces or lab data APIs where available and permitted

• Versioned datasets in institutional catalogs

• Notebook and experiment-tracking systems used in iNQspace

Exact integrations depend on cohort infrastructure and compliance.

Relationship to coursework¶

Courses that emphasize tool building or integration will specify required vs optional MCP engagement in the course syllabus. Program-level expectations live in ASSESSMENT.md and CURRICULUM_FULL.md.