API & MCP

The FastAPI backend publishes its full OpenAPI schema at /docs on any running instance. The essentials:

Jobs (the primary surface)

Long-running agent queries execute as durable jobs — never inside the HTTP request.

# enqueue (returns in milliseconds)
curl -X POST $BASE/api/jobs -H "Content-Type: application/json" \
  -d '{"type": "research", "query": "Which trials use pembrolizumab?"}'
# -> {"job_id": "..."}

# follow progress: SSE that REPLAYS the full event log, then live events.
# GET, so a browser EventSource gets automatic reconnection for free.
curl -N $BASE/api/jobs/{job_id}/stream

# or poll
curl $BASE/api/jobs/{job_id}

type is research (Smart Table), landscape, or catalysts. Events are status, progress, and terminal result/error — a reconnecting client rebuilds exact progress state from the replay.

Watchlist

curl $BASE/api/watchlist                          # list
curl -X POST $BASE/api/watchlist \
  -d '{"type": "company", "value": "Coherus BioSciences"}'
curl -X POST $BASE/api/watchlist/check            # diff both live sources now
curl $BASE/api/watchlist/digest                   # latest change digest

Exports

POST /api/export/excel and POST /api/export/pptx take a SmartTableResponse body (i.e. a job's result) and return the file — no agent re-run.

Synchronous endpoints

POST /api/research, /api/landscape, /api/catalysts still execute in-request — kept for CLI use and simple scripting. For anything user-facing, use jobs.

MCP server

mcp_server.py exposes ten tools over stdio to any MCP client:

search_clinical_trials · query_knowledge_graph · search_fda_records · search_fda_crls · search_pubmed_literature · search_sec_filings · search_corporate_news · get_safety_signals · get_exclusivity · run_smart_table

Each returns the same JSON observations the in-app agent consumes, including the has_results grounding flag. Configuration is in the Quickstart.