Fahali MCP server

Fahali — Financial Risk MCP Server for AI Agents

Give your agent a risk conscience. Fahali gives AI agents a callable financial-risk layer: portfolio risk, stress losses, crypto risk, contagion, crash precursors and market verdicts. Read-only — your agent already fetches prices; Fahali tells it what those prices mean for capital at risk. It returns a judged read with its confidence, the inputs it is missing, and a signed receipt, so your agent can check its own conviction against something scored against outcomes instead of raw prices it has to interpret blindly.

Observation, not advice. Fahali watches and explains market risk. It never tells you or your agent what to buy or sell. Read-only: no order routing, no path to capital.

Why an agent cites Fahali instead of a price feed

• Verified lead time. Per engine, the median hours ahead of the move on its correct material warnings, scored against what happened, with the misses in the same record. Your agent can weight a signal by proven lead, not by tone. Public, no key: GET /api/track-record/lead-time.
• Signed receipts. Every verdict carries a SHA-256 receipt and a provenance root. When your agent quotes Fahali, it can attach proof of what was said and when.
• A judged record with public replays. Predictive signals are graded against a disclosed, versioned method at their own horizon, and wrong calls stay on the record. Analytical engines are labelled observation-only rather than scored as forecasts, because grading an observation against a price move would manufacture a hit rate that means nothing. Methodology: fahaliai.com/methodology.
• Honest absence. Missing data says so. Quiet markets resolve as unresolved, never as wins. Some tools abstain by design: whale tracking returns an explicit "no on-chain data source" rather than inventing positions.

Coverage is crypto 24/7 and US equities and ETFs during market hours. Live coverage and engine counts are reported at GET /api/public/stats rather than asserted here, because the active set changes.

See it before you sign up

The lead-time record is public. This is the fastest way to confirm Fahali is scored against outcomes:

curl https://app.fahaliai.com/api/track-record/lead-time

Connect

Remote MCP (recommended, nothing to install):

Streamable HTTP:  https://mcp.fahaliai.com/mcp
SSE (legacy):     https://mcp.fahaliai.com/sse

Auth: OAuth 2.1 ("Connect with Fahali") or a static key, Authorization: Bearer sklive.... Grab a free developer key (50 calls/day, no card) at app.fahaliai.com/developer. In Claude, ChatGPT, or Cursor, add the URL above as a connector. Listed in the official MCP registry as com.fahaliai/fahali.

Run this repo as a local stdio proxy:

npm install
npm run build
FAHALI_API_KEY=sk_live_... node dist/index.js --stdio

It forwards each tool call to the upstream Fahali API; tier is enforced upstream.

Example: a pre-trade risk check

examples/pre-trade-check is a runnable agent (zero dependencies, one command) that asks Fahali, before a position is opened: is a flash-crash precursor firing, what is the judged read and what data is missing, and how early does the record say this engine usually warns. It then prints a PROCEED / CAUTION / HOLD posture and the receipt.

Tools

Around 30 tools. Most are read-only market intelligence; a few agent-workspace tools (memory, custom alerts) write only to your own workspace, never to markets. Highlights:

Shortened. The full README is on GitHub.