SEC Intelligence

🔎 SEC Intelligence MCP — Document-Grounded Financial AI

Ask Claude Desktop real questions about SEC filings — 10-Ks, 10-Qs, 8-Ks — and get answers quoted directly from the actual filing text, with a citation to the exact section (and page, where available) on every claim. Not a summary from training data. Not a guess.

Try it now: uvx sec-intelligence-mcp — see Quick install below.

Why this is different

Most finance-related MCP servers are data-API wrappers: they hand back numbers (revenue, EPS, price) from a database. None of the ones we surveyed read the actual filing documents, so none can answer a question that requires understanding what a company's management actually *said* — e.g. "how did NVIDIA's management explain the datacenter revenue surge?" or "did Amazon's forward guidance tone change between quarters?"

This server retrieves and quotes the real filing text with a citation on every claim, and its answer-generation prompt explicitly refuses to use prior/general knowledge when the retrieved passages don't contain the answer — verified live: asking about NVIDIA's non-existent "Mars operations" correctly returns "not present in the filing" rather than an invented answer. It also ships an automated RAGAS evaluation harness (see Evaluation results) that measures this claim on 50 real questions rather than just asserting it.

🚀 Quick install

Published on PyPI: https://pypi.org/project/sec-intelligence-mcp/. No clone, no build step — uv fetches and runs it on demand:

uvx sec-intelligence-mcp

That's the whole install. Two more things and you're ready to use it in Claude Desktop:

1. Get your free API keys

All four take under 5 minutes total, no credit card anywhere.

Don't have Docker or want to skip signing up for Qdrant? See the note in Getting Qdrant running below.

2. Connect Claude Desktop

Add this to your claude_desktop_config.json (Windows: %APPDATA%\Claude\claude_desktop_config.json, macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

{

"mcpServers": { "sec-intelligence-mcp": { "command": "uvx", "args": ["sec-intelligence-mcp"], "env": { "GEMINI_API_KEY": "your-key", "QDRANT_URL": "http://localhost:6333", "LANGFUSE_SECRET_KEY": "your-key", "LANGFUSE_PUBLIC_KEY": "your-key" } } } }

Restart Claude Desktop, open the tools list (hammer icon), and confirm sec-intelligence-mcp appears with a ping tool. That confirms the connection works before you rely on it for a real question.

3. Try it out

Every session starts by ingesting the company you want to ask about — that's what indexes its filings so they can be searched. Then just ask in plain English:

You: Ingest NVIDIA's last 2 annual filings

Claude *(calls ingest_company_filings)*: Indexed 2 filings for NVIDIA — 312 chunks

from FY2024's 10-K, 287 from FY2023's.

You: What did they say about risks from export controls?

Claude *(calls analyze_filing)*: NVIDIA's FY2024 10-K identifies export control

regulations as a primary risk: "The U.S. government has imposed, and may in the future

impose, controls on the export of our products... restrictions to China, Hong Kong, and

Russia have materially impacted our revenue." — *[Item 1A — Risk Factors]*

You: How does that compare to AMD?

Claude *(calls compare_companies)*: [grounded side-by-side answer, cited to each

company's own filing]

No prompt engineering, no special syntax — Claude picks the right tool automatically based on what you ask.

🧰 Available tools

🛠️ Tech stack

Open source and free-tier first — no paid API is required to run this end to end.

🏗️ Architecture

flowchart LR

A[Claude Desktop / MCP client] -->|MCP tool calls| B[sec-intelligence-mcp server] B --> C[SEC EDGAR API] B --> D[Qdrant vector search] B --> E[Gemini answer generation] B --> F[LangFuse tracing + eval scores] C -->|filings| B D -->|cited passages| B E -->|grounded answer| B

📊 Evaluation results

Measured with RAGAS on 50 hand-verified question/ground-truth pairs across 5 companies (full methodology and raw results in eval/README.md):

CI's eval-gate fails any PR to main that drops faithfulness below 0.75 on a real, live-ingested subset of these questions — see .github/workflows/ci.yml.

LangFuse dashboard

A real trace of analyze_filing answering "What risks does NVIDIA face from export controls?" — the span tree shows retrieval and embedding nested under the tool call, alongside the LLM generation, with a faithfulness: 1.00 score attached automatically:

👩💻 For developers

Want to run from source, contribute, or self-host instead of using the published package?

Getting Qdrant running

The simplest path is Docker: docker run -p 6333:6333 qdrant/qdrant. No Docker? Use a free Qdrant Cloud cluster instead and set QDRANT_API_KEY too.

Run from a local clone

1. Install uv.
2. Clone the repo and install dependencies:

   git clone https://github.com/jahanv01/sec-intelligence-mcp.git

cd sec-intelligence-mcp uv sync

3. Copy `.env.example` to `.env` and fill in the keys from the table above.

1. Start Qdrant locally:

   docker compose up -d qdrant

1. Run the server directly:

   uv run python src/server.py

Or with the MCP Inspector (dev UI, requires Node.js):

   uv run mcp dev src/server.py

For Claude Desktop, point it at your clone instead of the published package:

{

"mcpServers": { "sec-intelligence-mcp": { "command": "uv", "args": [ "--directory", "C:\\ABSOLUTE\\PATH\\TO\\sec-intelligence-mcp", "run", "python", "src/server.py" ] } } }

src/config.py fails fast at import time (raises RuntimeError) if any required key is missing.

Running via Docker

docker compose up -d builds the server image and starts it alongside Qdrant. The app service reads secrets from your local .env via env_file, and QDRANT_URL is overridden to http://qdrant:6333 (the in-network service name) since localhost inside the container would not reach the qdrant container.

Testing locally

uv run python -c "import mcp"                    # SDK installed correctly

uv run python scripts/test_server_stdio.py # server responds over stdio (ping -> pong) docker compose up -d qdrant uv run python scripts/test_qdrant.py # Qdrant round-trip works

EDGAR data layer (each hits the real EDGAR API)

uv run python scripts/test_edgar_lookup.py # ticker -> CIK, DuckDB-cached uv run python scripts/test_edgar_filings.py # recent 10-K filings for a ticker uv run python scripts/test_edgar_parser.py # download + clean a real filing uv run python scripts/test_edgar_sections.py # section detection + metadata

Embedding & retrieval pipeline (real model + real Qdrant)

uv run python scripts/test_chunker.py # section/paragraph chunking uv run python scripts/test_encoder.py # E5 embedding shape/latency uv run python scripts/test_ingest.py # chunk -> embed -> upsert to Qdrant uv run python scripts/test_search.py # semantic search with citations

MCP tools (real pipeline + real Gemini calls)

uv run python scripts/test_tool_ingest_company_filings.py uv run python scripts/test_tool_search_filings.py uv run python scripts/test_tool_analyze_filing.py uv run python scripts/test_tool_get_filing_summary.py

Full mocked suite (no network calls)

uv run pytest tests/ -m "not integration"

Project structure

src/

├─