The MCP server (openEHR for LLM agents)

The most forward-looking thing anarchie ships is anarchie mcp - a Model Context Protocol server that exposes the CDR to LLM agents. MCP is the emerging standard for giving language models typed, discoverable tools; wiring anarchie up to it means an assistant in Claude Desktop, or any agent framework, can read and write a real, validated openEHR record.

It is the same store and the same ops layer as the CLI and the REST API - just a third front end. It speaks JSON-RPC 2.0 over stdio (stdin/stdout), which is exactly what MCP clients launch and talk to.

Start it and list its tools

anarchie mcp reads JSON-RPC requests on stdin and writes responses on stdout. Here is a raw exchange - an initialize handshake followed by tools/list - piped straight in:

printf '%s\n%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/list","params":{}}' \
  | anarchie mcp

The initialize response identifies the server ({"name":"anarchie","version":"0.0.0"}), and tools/list returns the seven tools it offers:

Call a tool

tools/call invokes one by name. Listing the templates an agent can author against:

printf '%s\n%s\n' \
  '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' \
  '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"list_templates","arguments":{}}}' \
  | anarchie mcp

{
  "templates": [
    "adverse_reaction_list.v1",
    "problem_list.v1",
    "vital_signs_encounter.v1"
  ]
}

Why this is the interesting part: validation as a feedback loop

The compelling move is what happens on a bad commit. When commit_composition is handed a nonconformant Composition, it does not just fail - it returns the structured validation report in-band, marked isError: true, naming the exact openEHR path that breached and why (for example, a systolic magnitude outside the permitted range for mm[Hg]).

That turns the validator into a self-correction loop. An LLM asked to author a blood-pressure Composition can commit a draft, read back the precise violation, fix that field, and retry - the same machine-readable feedback that makes anarchie commit strict is what lets an agent converge on conformant openEHR data without a human in the loop.

agent: commit_composition(draft)      → isError: true, "…/value/magnitude outside range for mm[Hg]"
agent: (corrects the magnitude)
agent: commit_composition(fixed)      → committed, version_uid …::1