The MCP server — connect an AI agent to FedGrade

The FedGrade MCP server lets any Model-Context-Protocol-capable agent — Claude, ChatGPT, Copilot, or your own — query FedGrade's federal-civilian intelligence directly. Every factual payload a tool returns carries a provenance grade and citations. There are no ungraded answers, ever. The full reference lives at /docs/mcp; this is the orientation.

What it is

• Endpoint: https://fedgrade.com/api/mcp
• Transport: MCP Streamable HTTP (POST JSON-RPC). Stateless in v1 — each POST is an independent exchange, no sessions to manage.
• Auth: an Enterprise+ API key as a Bearer token for live mode, or no header for sample mode (a small static demo research library, free to try from any client).

The nine tools

• search_federal_corpus — hybrid retrieval over the research library, ranked + source-cited.
• get_entity — resolve a vendor/agency and return its graph profile.
• get_daily_brief — your persisted, citation-validated brief.
• verify_claim — run the deterministic claim-support validator on a (claim, excerpt) pair. Runs even in sample mode; touches no FedGrade data.
• get_provenance — the full audit trail behind one brief.
• assess_capture — the decomposed pWin score, recommendation, and graded factors.
• rank_opportunities — the ranked opportunity pipeline.
• query_entity_graph — walk an entity's relationship paths with the source behind each edge.
• get_forecast_calibration — FedGrade's published forecast track record.

The provenance promise

Every tool result — including error and not-found paths — carries a provenance_grade (verified / source-cited / modeled / insufficient-evidence) and a citations array. Agents should treat ungraded facts as nonexistent. The grade ladder is enforced in code and by a contract test that walks every tool — it is not marketing copy.

Connecting

The MCP docs carry copy-paste setup for Claude Desktop, Claude Code, ChatGPT, and raw cURL. In short: add the endpoint URL as a custom connector, and attach your fg_live_… key as the Authorization header for live data.

Limits

Tool calls share the platform query budget (30/min per key, per IP in sample mode). v1 is stateless, read-only, and returns JSON (no streaming partials). Write operations land behind the same provenance contract in a later version. See API keys for issuing the Bearer token and the REST API for the same data over HTTP.