FIG — Financial Intelligence Gateway

Trust & routing layer for financial agent services.

FIG is an MCP server that helps AI agents discover, verify, and compare 30+ financial data providers. It doesn't provide data itself — it tells agents which financial MCP server to trust and which best fits their task.

Think of it as Morningstar + Google for financial agents: rate the quality of financial services, and help agents find the best one.

The financial MCP ecosystem is exploding — Polygon.io, FactSet, Alpha Vantage, Alpaca, Bloomberg, and dozens more now offer agent-native APIs. But agents face a critical gap: who verifies the verifiers?

• An agent can't travel back in time to verify a signal source's track record
• Self-auditing is structurally worthless
• No existing service provides independent, verifiable trust scoring for financial MCP servers

FIG fills this gap with commit-reveal attestation: every judgment is SHA256-hashed before the outcome is known, published publicly, and independently verifiable by any agent.

git clone https://github.com/leoncuhk/fig.git
cd fig
uv sync

uv run fig
# or
uv run python -m fig.server

The server starts on http://127.0.0.1:8100/mcp (Streamable HTTP).

uv run pytest -v

# Commit today's regime judgments
uv run python -m fig.trust.daily_attest --phase commit

# Reveal yesterday's judgments with mark-to-market outcome
uv run python -m fig.trust.daily_attest --phase reveal

Payment via x402 (USDC on Base). Phase 0: all tools are free.

fig/
├── src/fig/
│   ├── server.py              ← MCP Server entry point (Streamable HTTP)
│   ├── config.py              ← Environment-driven configuration
│   ├── errors.py              ← Structured error types
│   ├── data.py                ← Market data via yfinance (public API)
│   ├── regime.py              ← Self-contained regime scoring (first-principles)
│   ├── evaluation.py          ← Strategy evaluation (bootstrap Sharpe CI)
│   ├── tools/                 ← MCP tool implementations
│   │   ├── discover.py        ← Financial service discovery + routing
│   │   ├── verify.py          ← Track record + attestation verification
│   │   ├── regime.py          ← Regime intelligence (self-contained)
│   │   └── evaluate.py        ← Strategy evaluation + signal verification
│   ├── trust/                 ← Trust spine
│   │   ├── attestation.py     ← Commit-reveal core logic
│   │   ├── daily_attest.py    ← Daily attestation cron
│   │   └── provider_scoring.py← Multi-dimensional provider trust scoring
│   ├── payment/               ← Payment layer
│   │   └── x402.py            ← x402 paidTool scaffold
│   └── registry/              ← Curated provider catalog
│       └── providers.json     ← 11 seed providers, manually reviewed
├── tests/                     ← 94 tests, full module coverage
├── daily/                     ← Commit-reveal attestation data
├── .well-known/               ← Agent discovery (A2A Agent Card)
└── docs/DESIGN.md             ← Full research + product definition

FIG's trust doesn't come from authority — it comes from verifiable history.

1. FIG computes a regime judgment for SPY
2. SHA256(canonical_json(judgment)) → commit hash published
3. Next day: judgment revealed with actual market outcome
4. Anyone can verify: recompute the hash, compare with committed hash
5. Over time: hundreds of independently verifiable daily judgments

Every judgment — including wrong ones — is published. Cherry-picking is structurally impossible because the commit hash locks in the judgment before the outcome.

FIG's first "customers" are agent builders (humans), not agents directly:

1. Human quants see the methodology + open track record → trust forms
2. Agent builders configure FIG as their agent's default financial MCP server
3. Agents inherit the builder's trust configuration
4. Later: agents with sufficient autonomy verify FIG's track record independently

All configuration is via environment variables:

FIG curates a registry of financial MCP servers with independent trust scoring across 5 dimensions:

Current registry includes 11 providers: Polygon.io, Alpha Vantage, Alpaca, FactSet, Bloomberg, Morningstar, S&P/Kensho, LSEG/Refinitiv, Financial Datasets, Trellis, and Collective2.

We publish our losers because that's how trust works. If we only showed winners, you couldn't trust the track record.

• Phase 0 (current): MCP server + daily attestation + free tools
• Phase 1: x402 paid tools + expanded provider scoring
• Phase 2: Provider self-certification + meta intelligence tools
• Phase 3: Builder subscriptions + priority routing

MIT