Financial Data

An MCP server (spec 2025-11-25) for SEC EDGAR + FRED + Polygon.io with OAuth 2.1, a citation-grounded 10-K extractor powered by Claude Sonnet 4.6, and an MCP Apps inline UI for the extracted summary.

Anchor audience: Anthropic Forward Deployed Engineering, Bridgewater / Citadel / Anthropic Finance teams.

CI Eval (nightly) Python License Release

About: MCP server for SEC EDGAR + FRED + Polygon with OAuth 2.1, a citation-grounded 10-K extractor, and MCP Apps inline UI — built for Anthropic FDE, Bridgewater / Citadel quant, and Cursor FDE reviewers.

Demo

Start with the interactive deck for the full walkthrough, or scroll the static slide exports below.

Step	Command
Present	docs/presentation.html — 14-slide interactive project deck (open in any browser; ← → to navigate, O for overview). PowerPoint version: docs/mcp-financial-data-presentation.pptx
Live MCP	docs/demo/cursor-mcp-setup.md — optional Cursor appendix
Regenerate slides	make presentation-export — re-export PNGs from the HTML deck

Story: analyst asks for AAPL Item 1A risks → cited TenKSummaryCard → eval proof at 1.0 → stack + CI.

All 14 presentation slides (static PNG exports from docs/presentation.html)

2 — The problem

3 — What it is

4 — Architecture

5 — The citation contract

6 — The output card

7 — OAuth 2.1 security

8 — Data discipline

9 — Eval harness

10 — Engineering rigor

11 — The stack

12 — Story demo

13 — Results (v0.1.0)

14 — Recap & close

Why this exists

Financial-services AI work consistently fails the same audit checklist:

1. The agent answered with a number, but the citation pointed at the wrong filing.
2. The agent dropped a fact silently when the source didn't support it.
3. The MCP integration didn't validate JWT scopes per request.
4. The eval harness wasn't deterministic, so the regression yesterday is indistinguishable from a flaky judge today.

Every one of these is a hard constraint in this repo, enforced in CI.

Architecture

Every request crosses the OAuth 2.1 boundary before anything runs; every external fetch is rate-limited and cached; every tenk.* claim is citation-grounded before it leaves the server. Thick arrows are the live request path, dotted arrows are control or replay paths.

flowchart LR
    C["MCP Client<br/>Claude Desktop · Cursor · Goose"]
    OAuth["OAuth 2.1 Resource Server<br/>JWT validation · alg allowlist · PKCE"]

    subgraph core["FastMCP Server — MCP spec 2025-11-25"]
        Router["Tool router +<br/>MCP Apps registry"]
        EDGAR["edgar.list_filings<br/>edgar.company_facts"]
        FRED["fred.series"]
        Polygon["polygon.aggregates"]
        Extractor["tenk.extract_section<br/>citation-grounded 10-K"]
        UI["TenKSummaryCard<br/>inline UI · citation pills"]
        Evals["Eval harness<br/>smoke · full · offline"]
    end

    Cache[("SQLite cache · 24h TTL")]
    Anthropic["Anthropic Messages API"]

    C ==>|"Bearer JWT"| OAuth
    OAuth -.->|"401 + WWW-Authenticate"| C
    OAuth ==>|"authenticated"| Router
    Router --> EDGAR
    Router --> FRED
    Router --> Polygon
    Router --> Extractor
    Extractor --> UI
    Extractor ==>|"Sonnet 4.6 + Citations API"| Anthropic
    Evals ==>|"Opus 4.7 judge"| Anthropic
    Evals -.->|"replays tool calls"| Router
    EDGAR --> Cache
    FRED --> Cache
    Polygon --> Cache
    Extractor --> Cache

    classDef client fill:#15324f,stroke:#4f8cff,color:#eaf1ff
    classDef gate fill:#3a2c12,stroke:#ffb347,color:#ffe9c7
    classDef tool fill:#16233a,stroke:#2d3a4f,color:#dce6f5
    classDef key fill:#1d2f4d,stroke:#4f8cff,color:#eaf1ff
    classDef store fill:#16301f,stroke:#2ee08a,color:#d6f5e3
    classDef ext fill:#271d3d,stroke:#9b86ff,color:#ece6ff

    class C client
    class OAuth gate
    class Router,EDGAR,FRED,Polygon,UI tool
    class Extractor,Evals key
    class Cache store
    class Anthropic ext

Node style	Meaning
Amber gate	OAuth 2.1 boundary — validated on every request
Blue (filled)	Citation-grounded extractor + eval harness — the audited paths
Green store	24-hour SQLite response cache — keeps eval runs reproducible
Purple	External Anthropic Messages API (Sonnet 4.6 extract, Opus 4.7 judge)

Quickstart

git clone https://github.com/SebAustin/mcp-financial-data.git
cd mcp-financial-data
cp .env.example .env   # fill in API keys
make setup             # uv sync + pre-commit install
make ci                # lint + typecheck + tests + smoke eval (offline)
make serve             # run the MCP server on $MCP_HOST:$MCP_PORT

What's in the box

Surface	Tool / endpoint	Notes
MCP tool	edgar.list_filings	Recent SEC filings for a CIK.
MCP tool	edgar.company_facts	XBRL facts (us-gaap concepts).
MCP tool	fred.series	FRED economic time series.
MCP tool	polygon.aggregates	OHLCV aggregate bars.
MCP tool	tenk.extract_section	Claude-grounded 10-K claims with citations.
MCP App	tenk-summary-card	Inline UI rendering the extractor output.
OAuth	RFC 6750 resource server	Validates JWT against external IdP.

Eval targets (W1)

The eval harness writes per-case JSONL plus a summary JSON to evals/runs/<run_id>/. CI runs --smoke --offline on every PR; nightly CI runs --full --budget 5 --min-judge-score 0.85 against live APIs.

Metric	W1 target	W1 actual	Source of truth	Notes
Smoke pass rate	5 / 5	5 / 5	evals/cases/seed.jsonl	Offline fixtures.
Mean exec-accuracy	≥ 0.95	1.00 (offline) / 0.80 (live)	evals/metrics.py::exec_accuracy	Live MSFT XBRL multi-row list match — tracked in follow-on issues.
Mean citation coverage	= 1.00	1.00	evals/metrics.py::citation_coverage	Required for tenk.*.
Mean judge score (offline)	≥ 0.90	1.00	evals/metrics.py::judge_with_stub	0.5·exec + 0.5·citation.
Mean judge score (live full)	≥ 0.85	0.93	evals/judge.py::judge_with_claude	Opus 4.7 five-axis rubric.
P50 latency (smoke)	≤ 50 ms	< 1 ms	harness latency_ms	Offline only.
Total cost / smoke run	$0.00	$0.00	harness total_cost_usd	--offline enforced.
Total cost / live full run	≤ $5.00	$0.08	harness total_cost_usd	--budget 5 gate.
Coverage gate (src/)	≥ 85%	~87%	pytest --cov-fail-under=85	mypy --strict also gates.

Latest --full runs (git b8a4ba3, 2026-05-21)

Reproduce offline:

uv run python -m mcp_financial_data.evals.harness --full --offline

Reproduce live (requires .env secrets; ~$0.08 per run):

uv run python -m mcp_financial_data.evals.harness --full --budget 5 --min-judge-score 0.85

Run	run_id	Pass	mean exec-acc	mean citation	mean judge	cost USD
Offline full	20260521T114816Z_b8a4ba3	5 / 5	1.00	1.00	1.00	0.00
Live full	20260521T114826Z_b8a4ba3	5 / 5	0.80	1.00	0.93	0.08

See CHANGELOG.md for the v0.1.0 release notes.

Hard constraints (skim before contributing)

• Citations are non-optional for the 10-K extractor. Every CitedClaim.text carries at least one Citation. Uncited model output is dropped or moved to notes with [INFERENCE]. See docs/adr/0003-citations-required-for-extracted-claims.md.
• EDGAR Fair Access is enforced. Every request carries the EDGAR_USER_AGENT env value. Process rate limit ≤ 10 req/sec.
• Spend cap. MAX_API_SPEND_USD defaults to 50. Enforced in extractor and harness.
• No requests, no print(), no subprocess shell=True, no bare except.
• OAuth 2.1 RS only. This server validates JWTs; it is never the IdP. See docs/adr/0004-oauth21-as-resource-server.md.

Layout

src/mcp_financial_data/      # the package
  server.py                  # FastMCP entrypoint
  auth/oauth.py              # OAuth 2.1 RS primitives
  tools/{edgar,fred,polygon} # async API clients
  extractors/tenk.py         # citation-grounded 10-K extractor
  apps/ui.py                 # MCP Apps inline UI registration
  evals/{harness,metrics}    # eval harness (--smoke / --full / --offline)
tests/{unit,integration}     # pytest, 85% gate, integration skipped by default
evals/cases/seed.jsonl       # 5 hand-authored eval cases
evals/runs/                  # per-SHA harness output (gitignored)
docs/adr/                    # MADR architecture decisions
.github/                     # CI templates + issue/PR templates + dependabot

References

1. Model Context Protocol — specification 2025-11-25
2. SEC EDGAR — accessing data fairly (User-Agent + 10/sec)
3. FRED API documentation
4. Polygon.io REST API
5. Anthropic Citations API

License

MIT. See LICENSE.