Normalized US-GAAP financial statements & fundamentals from SEC EDGAR, as MCP tools for LLM agents.
MCP Server: ai.datavidence/datavidence-financials
The MCP server ai.datavidence/datavidence-financials provides normalized US-GAAP financial statements and fundamentals from SEC EDGAR. It exposes this content as MCP tools for LLM agents, enabling automated retrieval and use of financial and fundamental data derived from EDGAR filings.
๐ ๏ธ Key Features
Normalized US-GAAP financial statements
Fundamentals data
Source: SEC EDGAR
Delivered as MCP tools for LLM agents
๐ Use Cases
Retrieve normalized US-GAAP financial statements for analysis
Access EDGAR-derived fundamentals for agent workflows
โก Developer Benefits
Standardized, normalized financial data suitable for LLM agent use
Tool-based access to SEC EDGAR-derived statements and fundamentals
โ ๏ธ Limitations
Scope is limited to normalized US-GAAP financial statements & fundamentals from SEC EDGAR
Normalized US-GAAP financial statements for every SEC XBRL filer โ point-in-time, as originally reported โ as MCP tools for your AI agent (or a plain REST API).
Language models are good at reasoning about financial statements and bad at getting them. SEC EDGAR has the data, but its XBRL is a thicket of inconsistent tags, renamed concepts, and per-filer quirks โ so agents either hallucinate the numbers or you spend weeks building a normalization layer. This is that layer: your agent asks for a company's financials and gets a clean, consistent answer โ over the Model Context Protocol, or as a plain REST API.
โญ If this is useful, please star the repo. It's a solo, early project, and stars are how other developers find it.
Claude calls the Datavidence Financials get_revisions tool: Pfizer's FY2019 revenue as filed ($51.75B) and as it reads today ($40.9B), with links to the SEC filings
This is the open-source, AI-native layer for the Datavidence Financials API โ a standalone stdio server that forwards over HTTPS to the REST API (bring your own key). The connector is MIT-licensed; the underlying API is a commercial service.
Try it in 30 seconds (no signup)
Run it with no clone using uv โ it installs into a throwaway environment:
sandbox_demo_key returns sample data with no signup โ good for a first run. Set FL_API_KEY to your real key for live SEC data. The server speaks MCP over stdio, so an MCP client (e.g. Claude Desktop, below) launches it.
For the sample data, use the /mcp URL and leave out the header.
claude.ai: add https://mcp.financials.datavidence.ai/mcp as a custom connector with no sign-in. claude.ai can't send an API key for most accounts yet, so it gets the sample data for now.
Why it's different
Point-in-time, as originally reported. Pass an as_of date and you get the figures as they stood then โ no look-ahead bias from later restatements. The thing backtests quietly get burned without.
Normalized to one schema. EDGAR's XBRL tags drift between filers and change over the years. We map them to a single US-GAAP income statement, balance sheet, and cash flow โ including the awkward cases (noncontrolling interest, mezzanine/temporary equity, non-calendar fiscal years) โ so you never parse a filing.
Every SEC XBRL filer. Coverage is the whole SEC XBRL universe (~2009 to present), fetched and normalized on demand โ any CIK with XBRL facts, not a curated subset.
Source-linked, too. Turn on provenance and every value carries its SEC accession, filed date, and a direct EDGAR URL โ so the model cites the filing instead of inventing a number.
Honest scope: US-GAAP only (IFRS filers return a clear "unsupported" signal rather than wrong numbers); XBRL-era history (~2009 onward). And when a company reorganizes under a new holding entity, SEC's ticker map points only at the successor CIK โ the earlier years stay under the predecessor, so the error tells you and you pass cik= directly.
Tools
get_financials โ normalized income statement + balance sheet + cash flow
for one company and fiscal year, from SEC EDGAR XBRL. Supports as_of
(point-in-time, as originally reported), include_provenance (per-value SEC
accession, filed date, and direct EDGAR source URL), and include_ratios
(margins, ROA/ROE, leverage).
get_financials_batch โ the same, for up to 25 companies in one call
(comma-separated tickers/CIKs). Per-symbol results; counts as a single quota
unit, so one bad symbol never fails the batch.
list_filings โ a company's recent SEC filings (newest first) from the
EDGAR submissions index; filter by form (e.g. 10-K). Use it to discover which
fiscal years are available before calling get_financials.
get_revisions โ how a fiscal year's figures CHANGED across filings: every
reported value with its filing date, form, SEC accession and EDGAR link, plus
whether it was restated and by how much. Pfizer's FY2019 revenue was filed at
$51.75B in 2020 and restated to $40.9B by 2022 โ this is how you see that.
search_companies โ find a ticker and CIK by ticker or company name, for
when you have "Berkshire" and need BRK-B. Use it before the tools above,
which all require an identifier.
get_usage โ the key's monthly quota (tier, used, remaining, reset). Free
to call; consumes no quota.
Errors surface the API's recovery_action hint, so agents self-correct.
Claude Desktop
Add to claude_desktop_config.json, then fully quit (Cmd+Q) and reopen Claude
Desktop โ the tools appear:
Then ask, e.g., "Get AAPL's FY2023 income statement, balance sheet, and cash flow with ratios." Swap in your real key for live data.
Codex
Add to ~/.codex/config.toml, then start codex โ the tools appear:
toml
[mcp_servers.datavidence-financials]command = "uvx"args = ["datavidence-financials"]
env = { FL_API_KEY = "sandbox_demo_key" }
# uvx downloads the package on first run; allow more than Codex's 10 s default.startup_timeout_sec = 30
Swap in your real key for live data.
Or connect by URL instead (nothing to install; set DATAVIDENCE_API_KEY to your RapidAPI key):
Leave out bearer_token_env_var and use https://mcp.financials.datavidence.ai/mcp for the sample data.
Configuration (environment variables)
Variable
Default
Notes
FL_API_KEY
(none)
Your API key. sandbox_demo_key = no-signup trial (sample data).
FL_API_BASE_URL
https://api.financials.datavidence.ai
Point at a marketplace gateway for metered BYOK access.
FL_API_KEY_HEADER
X-API-Key
Header to send the key in (e.g. X-RapidAPI-Key for the RapidAPI gateway).
FL_API_TIMEOUT
30
Per-request timeout, in seconds.
Getting live data
The demo key returns sample data. For live figures, grab a key on RapidAPI: Free (1,500 calls/mo), Pro ($29 / 50k), Business ($99 / 300k). Full feature parity across tiers โ they differ only by monthly call volume.
To use your RapidAPI key, set these three variables (in the env block of the Claude Desktop config, or the env table in Codex's config.toml):
Your key is on RapidAPI under your app's Authorization page, or in any code snippet on the API's page. Calls are metered by RapidAPI against your plan. With a RapidAPI key, get_usage shows your RapidAPI plan and its monthly limit; RapidAPI meters your calls, so check how many you've used in your RapidAPI dashboard.
29 tests, no network and no API key needed. They call every tool through the MCP SDK
with the HTTP layer stubbed, check that an API error reaches the model with its recovery
hint, and check that each tool declares all four behaviour hints (read-only,
non-destructive, idempotent, open-world).
Privacy Policy
The connector runs on your machine, has no telemetry and stores nothing. When a tool is
called it sends two things to the Datavidence Financials API: the tool's arguments (for
example a ticker, a CIK, a fiscal year or a company-name search) and your API key, in a
request header. It does not see your files or the rest of your conversation.
Where it goes:https://api.financials.datavidence.ai by default. If you set
FL_API_BASE_URL to the RapidAPI gateway, requests go through RapidAPI, under its own
privacy policy.
What the API records: for each request, the endpoint, HTTP status, response time,
timestamp and IP address, tied to a hash of the API key (never the key itself).
Retention: usage logs are kept for 12 months.
Sharing: not sold; shared only with the marketplace gateway you subscribed through and
the infrastructure provider that hosts the service.
Found a company where the numbers come back wrong, or a field/tool you need? Open an issue โ the normalization edge cases are exactly the feedback that makes this better, and I read every one.