OKF MCP retrieval prototype

For end-user AI access, start with the access chooser. This adapter has no standard initialize handler or tool input schemas and is not ready for installation in an arbitrary MCP client. It reads one configured corpus from a corpora mapping with nodes and relationships; it does not load large descriptors or top-level edges. Its five operations are okf.list_bundles, okf.search, okf.get_record, okf.follow_relationships and okf.context_pack.

This prototype provides bounded, read-only access to a local okf-bundle.json. It is an empirical retrieval surface, not a claim of production deployment or certification against every MCP client revision.

Run the newline-delimited JSON-RPC adapter from the repository root:

uv run --locked python -m mcp.okf_mcp_server --bundle okf-bundle.json

Example request:

{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"okf.context_pack","arguments":{"question":"Why is YAML-LD additive to OKF 0.2?","limit":3,"max_bytes":12000}}}

The transport adapter exposes tool, resource and prompt-shaped JSON-RPC methods. The tested contract is the Python retrieval core: exact record IDs, deterministic lexical search, explicit relationship traversal, SHA-256 bundle identity and hard byte limits. A production server should wrap this core in the official SDK and the MCP revision supported by its target clients, then add authentication, authorisation, structured audit and operational monitoring.