MCP server

LOKF ships a Model Context Protocol server so an agent can drive the toolkit directly — load a bundle, run SPARQL, project RDF, propose relations — without shelling out to the CLI or parsing its text output. Every tool returns JSON-serializable values.

Running it

The server speaks MCP over stdio. Two equivalent ways to launch it:

uv run lokf mcp     # the CLI subcommand
lokf-mcp            # the dedicated console script (same server)

lokf-mcp is the entry point you point an MCP client at, because a client launches the server as a subprocess and talks to it over stdin/stdout.

Client configuration

Most MCP clients read an mcpServers block. Register lokf-mcp as the command:

{
  "mcpServers": {
    "lokf": {
      "command": "lokf-mcp"
    }
  }
}

If lokf is installed in a project virtualenv rather than on the global PATH, run it through uv instead:

{
  "mcpServers": {
    "lokf": {
      "command": "uv",
      "args": ["run", "lokf-mcp"]
    }
  }
}

The tools

The server exposes seven tools, mirroring the toolkit surface. Each returns JSON-serializable data (and a {"error": …} object on bad input rather than throwing):

Tool	What it does
list_concepts	Fast index of a bundle: {concept_id, type, title, iri} for every concept.
describe_concept	Full record of one concept — {iri, type, title, body, frontmatter, turtle}.
sparql_query	Run SPARQL over the bundle. SELECT → {columns, rows}, ASK → {boolean}, CONSTRUCT/DESCRIBE → {turtle}. Schema prefixes preset.
convert	Project a concept file or bundle directory to RDF: {format, rdf} (ttl, nt, jsonld, …).
propose_relations	Suggest typed relations from prose links; optionally apply them, with a per-proposal applied flag.
get_vocabulary	The schema vocabulary: {classes, relations} read from lokf.yaml.
bundle_summary	Orientation at a glance: {concept_count, types, triple_count, relation_edge_count}.

Each tool takes a bundle or concept path (and, where relevant, a SPARQL string or output format) and hands back structured data the agent can reason over.

What agents can automate

With the server connected, an agent can run an end-to-end loop over a knowledge base with no bespoke code:

• Orient — call bundle_summary and list_concepts to learn what a knowledge base holds before touching it.
• Author-and-check — write a concept, convert it (or read its turtle from describe_concept), and sparql_query the graph to confirm the new triples landed on the right subjects.
• Enrich — call propose_relations to surface typed edges the prose already implies, review the confidence and rationale, and apply the ones worth keeping.
• Answer questions — turn a natural-language question into a sparql_query against the bundle and read back the rows.
• Stay schema-correct — call get_vocabulary to discover which relations a class is allowed to use.

For the skills that package these workflows for coding agents, see Agents.