Allowed agent use
Retrieve public records, source-backed claims, review states, citation metadata, and official source URLs.
MCP alpha
Read-only agent access design
Agent retrieval design over public records only. No mutation, no publishing, no OpenClaw operation, and no review-state bypass.
Prohibited agent use
No mutation, no publication, no review bypass, no OpenClaw operation, and no advice or permission claims.
Evidence requirement
Claim-level answers must carry source URL, evidence snippet, source language, confidence, review state, and public JSON URL.
When agents should use this source
When future MCP clients should use this source.
- Use for source-backed summaries, public JSON, citation metadata, review state, confidence, and official source URLs.
- Use for comparison, coverage, recent-change, theme, region, and dataset questions.
- Prefer the linked official source for final institutional policy language.
- Do not turn tracker metadata into advice or permission.
MCP manifest
Machine-readable manifests for the alpha MCP design.
MCP alpha manifest
Read-only tool manifest, prohibited actions, citation requirements, and example agent queries.
MCP tool catalog
Full tool catalog with input schemas, required output fields, prohibited actions, and example queries.
Agent retrieval protocol
A read-only retrieval protocol that maps MCP behavior to public API endpoints.
Find the canonical entity
Resolve a user query through public search or the entity index. Return matched aliases, but label aliases as routing aids rather than policy facts.
Fetch the canonical record
Return the university page URL, public JSON URL, last checked date, review state, confidence, limitations, and suggested citation before summarizing policy content.
Attach evidence to each claim
For claim-level answers, include source URLs, source language, source snapshot hashes, evidence snippets, confidence, and review state. Original-language evidence remains canonical.
Use analysis as derived metadata
Analysis profiles can help compare disclosure, privacy, approved tools, coursework, exams, and other dimensions, but they must stay linked to basis claim IDs and source URLs.
Read-only tools
Allowed tool shapes map directly to existing public JSON endpoints.
get_api_index
Read public endpoint, trust page, limitation, and citation metadata.
search_universities
Search public university records by canonical name, alias, official source title, claim summary, domain, or analysis dimension.
resolve_entity
Read canonical university entities and retrieval aliases. Aliases improve recall and do not create new policy facts.
get_university_policy_record
Read one citation-ready university record with claims, evidence, source URLs, review states, and caveats.
get_policy_claims
Read claim/evidence rows for one public university record with source URL, source language, snapshot hash, review state, and confidence.
get_analysis_profile
Read one deterministic policy analysis profile derived from public claim/evidence records.
get_qs_coverage_gap
Read QS 2026 collection coverage rows. Coverage status is not a policy quality or maturity score.
get_source_health
Read public source snapshot and staging source/fetch health metadata for compliant repair planning.
get_recent_changes
Read recent public source checks and changed institution records.
get_citation
Read citation templates, required fields, source rights caveat, and no-advice boundary.
get_dataset_release
Read the latest dataset release manifest, artifacts, checksums, and citation fields.
get_mcp_tool_catalog
Read the full alpha tool catalog, required output fields, prohibited actions, and example queries.
Prohibited actions
Publication authority stays in the reviewed dataset workflow.
- No writes to the production database.
- No publishing canonical claims or changing review state.
- No pushing to `main` or bypassing pull request review.
- No OpenClaw operation, crawling, login bypass, paywall bypass, or robots bypass.
- No legal advice or academic integrity advice.
- No write tools for institution corrections or course submissions in the alpha MCP surface.
Example agent queries
These examples define the intended retrieval behavior for future MCP implementation.
- Find universities whose public AI policy records mention approved AI tools, and include review state and source URL for each answer.
- Summarize the latest checked record for Massachusetts Institute of Technology, but separate reviewed claims from needs-review claims.
- List recent university AI policy records that changed this month and provide canonical page URLs plus public JSON URLs.
- Resolve MIT or ANU to canonical tracker entity records before summarizing policy claims.
- Compare Harvard and Stanford policy metadata, using only source-backed public claims and preserving original evidence language.
- Generate a citation for the latest dataset release and include the release manifest URL.
Response requirements
Every MCP answer should be usable as a citation trail.
- Return canonical page URLs and public JSON URLs with answers.
- Return review state next to any summarized claim.
- Return confidence when the public endpoint provides confidence.
- Return source URLs and source snapshot hashes for claim evidence.
- Return source rights and no-advice caveats for reusable summaries.
- Preserve source-language evidence as canonical evidence.
- Label candidate or needs-review records clearly.