CogitaveLearn
Training
View as
MarkdownJSON

Summary

You now know a canonical model's native MCP interface - Cogitave's Core here - well enough to reason about a call to it, not just recite that one exists.

In this module, you:

  • Learned why MCP can be a canonical model's canonical query surface rather than an adapter: Cogitave's ADR-0003 chose tools/resources over MCP as the source of truth, with REST, GraphQL, and llms.txt generated from the identical catalog, graph, and search materializations, so humans and agents cannot drift onto different data.
  • Pinned the protocol baseline such a server commits to: spec revision 2025-11-25, stdio and Streamable HTTP transports, JSON-RPC 2.0 over a stateful session, JSON Schema 2020-12 for every tool's input and output, and tool input-validation failures returned as tool execution errors so a calling model can self-correct - and saw why the revision is pinned and dated rather than followed, since 2026-07-28 has since removed the session and the handshake entirely, and this site's own MCP endpoint already runs it.
  • Named the tools and resources such a surface exposes - docs_search through describe_schema, the cogitave://{type}/{id} resource template with subscribe and change notifications - and saw how query_graph is bounded (read-only, allowlisted labels, capped depth, capped rows) so it cannot become a general escape hatch.
  • Distinguished the open reads from the propose-only writes: request_intake and advance_stage stage a draft node and open a GitHub issue/PR, and never mutate protected state directly.

Next steps

  • The mcp-interface.md doc, in the estate's core repository, is worth a re-read with this module's context - it is short, and now every section should be legible on its own.
  • ADR-0003 is the estate's decision record itself; revisit it if you want the full trade-off behind "native."
  • Build on a single canonical model - this module is part 3 of 4 in that path; it also covers the property-graph model itself, the query interface (Cogitave Query is its reference implementation), and how products project their state into the canonical model.