JSON-RPC API

The HTTP control plane (daedalus.api) exposes the in-process training API over a single JSON-RPC 2.0 endpoint. It is a thin FastAPI app (src/daedalus/api/app.py) whose handlers dispatch into the shared METHOD_HANDLERS table in src/daedalus/api/methods.py — the same table the MCP server uses, so the two surfaces can never drift.

Endpoints

Method	Path	Purpose
POST	/rpc	JSON-RPC 2.0 — single request or a batch array
GET	/healthz	Liveness check — returns {"status": "ok"}

The app intentionally disables the OpenAPI docs UIs (docs_url=None, redoc_url=None): the JSON-RPC contract is the interface.

Starting the server

uv run daeda serve-api --host 127.0.0.1 --port 8000

Both flags are optional (defaults shown). The server runs under uvicorn; the /rpc handler is synchronous and Starlette runs it in a threadpool, so the blocking in-process / network JSON-RPC handlers never stall the event loop.

No auth yet

The server has no authentication or authorization. Bind it to a trusted network boundary only (127.0.0.1 by default); do not expose it publicly as-is.

Methods

The method names come straight from METHOD_HANDLERS:

Method	Params	Returns
catalog.list_views	(none)	One row per feature view (name, entity, join keys, feature count, source)
catalog.show_view	name	A view's schema: timestamp field, source, fields with dtypes
service.list	(none)	One row per feature service (name, version, owner, grain, column count, serving flag)
service.show	name	A service's full column contract (kind, dtype, default, nullable)
pipeline.compile	service	The compiled operator pipeline: operators + a graph of nodes/edges
runs.materialize_partition	service, partition	Triggers a daily training partition; returns {run_id, status}
runs.get	run_id	Run status, partition, start/end times, metadata
runs.list	service?, limit?	Recent runs (default limit 20)
runs.logs	run_id, tail?	Run log lines
assets.lineage	(none)	Asset lineage graph (nodes + edges)

The catalog.*, service.*, and pipeline.compile methods resolve and compile in-process. The runs.* and assets.lineage methods delegate to a Dagster runs backend: a DagsterGraphQLRunsBackend when DAGSTER_GRAPHQL_URL (or ApiContext.dagster_graphql_url) is set, otherwise an ephemeral in-process InProcessDagsterRunsBackend. See Dagster Orchestration.

Example request

curl -s http://127.0.0.1:8000/rpc \
  -H 'Content-Type: application/json' \
  -d '{
        "jsonrpc": "2.0",
        "method": "catalog.list_views",
        "params": {},
        "id": 1
      }'

{
  "jsonrpc": "2.0",
  "result": [
    {
      "name": "user_likes",
      "entity": "user",
      "join_keys": ["user_id"],
      "n_features": 6,
      "source": "likes"
    }
  ],
  "id": 1
}

Compiling a service to its operator DAG:

curl -s http://127.0.0.1:8000/rpc \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","method":"pipeline.compile","params":{"service":"dssm_ranking"},"id":2}'

A request with id omitted is a JSON-RPC notification: it is processed but produces no response body (HTTP 204 No Content).

Error responses

Errors use standard JSON-RPC error objects, {code, message, data?}:

Code	Meaning	Raised when
-32600	Invalid Request	Payload is not a valid JSON-RPC request object
-32601	Method not found	method is not in METHOD_HANDLERS
-32602	Invalid params	Missing/invalid params, or an unknown view/service/run
-32000	Internal error	Unexpected handler failure (server-side)

{
  "jsonrpc": "2.0",
  "error": {
    "code": -32601,
    "message": "Method not found",
    "data": { "method": "catalog.no_such_method" }
  },
  "id": 3
}

CORS

Cross-origin access is controlled by the DAEDALUS_API_CORS_ORIGINS environment variable — a comma-separated allow-list. It defaults to * (permissive, for local development); lock it down in production:

DAEDALUS_API_CORS_ORIGINS="https://ui.example.com,https://admin.example.com" \
  uv run daeda serve-api --host 0.0.0.0 --port 8000

The GUI planned in a separate stack consumes exactly this contract — see Platform overview.

See also

• MCP Server — the same methods as agent tools over stdio.
• Dagster Orchestration — the runs backend behind runs.*.
• CLI reference.