Skip to main content
Glama
open-finance-mcp

open-finance-mcp

by nathanaelhub

open-finance-mcp

An MCP server that gives Claude free, citable financial data for comps and DCF work: standardized financials from SEC EDGAR XBRL filings, filing links, Treasury yields from FRED, and market prices and beta. Every number comes back with its source: the XBRL concept, the filing accession and a URL to the filing.

Why

Anthropic's financial-services plugins ship /comps, /dcf and earnings workflows whose skills tell Claude to pull data from MCP connectors first and to cite every hard-coded input. All of the bundled connectors are paid terminals (FactSet, S&P Capital IQ, PitchBook, Morningstar, …). Without a subscription, the skills fall back to web search, which they explicitly warn against.

This server fills that gap with public sources, and it is built around the skills' citation requirement instead of treating sources as an afterthought.

Related MCP server: edgar-mcp-server

Tools

Tool

Returns

lookup_company

Ticker, name, CIK and exchange for a ticker or name search, including registrants with no current ticker

get_financials

Annual + LTM income statement, cash flow and balance sheet; derived EBITDA, FCF, total debt, net debt; per-value citations

get_filings

Recent 10-K / 10-Q / 8-K (etc.) with document and index URLs

get_earnings_release

Text of the latest (or an earlier) earnings press release from its Item 2.02 8-K, tables kept as rows, paged

get_market_data

Price, cover-page shares outstanding, market cap, 5-year monthly beta vs the S&P 500

get_treasury_yield

Latest constant-maturity Treasury yield (3M–30Y) from FRED

get_fx_rate

US dollars per unit of 18 currencies from FRED's H.10 rates: latest, on a date, or a period average

get_comps

Peer table: market cap, EV, LTM revenue/EBITDA/net income, EV/Revenue, EV/EBITDA, P/E, with median and mean

All tools are read-only. Nothing needs an API key.

Example: get_comps(["KO", "PEP", "KDP", "MNST"]), run 2026-09-29

Ticker

Period

Mkt cap ($B)

EV ($B)

EV/Rev

EV/EBITDA

P/E

Beta

KO

LTM to 2026-04-03

374

407

8.26

26.23

27.27

0.34

PEP

LTM to 2026-06-13

176

226

2.33

12.57

16.81

0.35

KDP

LTM to 2026-06-30

42

71

3.52

18.51

29.59

0.40

MNST

LTM to 2026-06-30

41

—

—

—

19.22

0.52

The notes that came back with those rows are the point of the design:

  • KO: EDGAR lists a 10-Q filed 2026-07-29 that the XBRL API does not include yet, so the figures stop at Q1. The tool says so instead of passing stale LTM off as current.

  • KDP: non-operating items of −1.3B exceed 25% of operating income, so the P/E is distorted and EV/EBITDA should carry more weight.

  • MNST: no debt concepts are reported, so EV is left null rather than assuming zero debt. A separately labeled enterprise_value_if_debt_free (37B) is offered, to use once the balance sheet confirms it.

Does it help? (evals)

Ten claude plugin eval cases, each run with and without the plugin (both arms can use web search), on live data, transcripts audited:

With plugin

Without

Mean score (10 cases, 3 runs each)

0.98

0.74

Latest quarter (Coca-Cola Q2 2026)

1.00

0.07

Cost per run (mean / peer comps)

$0.13 / $0.18

$0.52 / $2.96

The clearest wins are fresh data, cost and consistency. When the baseline does search the web, it mostly matches the plugin on judgment questions (bank multiples, DCF inputs), and an earlier, larger claimed gap there was withdrawn after transcripts showed it came from runs that never searched. Per-case results, the audit and the caveats are in plugin/evals/RESULTS.md.

Install

Requires uv. SEC requires automated clients to send a User-Agent with a contact, so you provide your name and email once.

As a Claude Code plugin (adds the server plus a skill on citing the data):

claude plugin marketplace add nathanaelhub/open-finance-mcp
claude plugin install open-finance@open-finance
claude plugin configure open-finance   # set sec_user_agent: "Your Name you@example.com"

If the contact is missing, the server still starts and its tools answer with an error saying how to set it, rather than silently not loading. It can also come from the OPEN_FINANCE_SEC_USER_AGENT environment variable.

It works alongside financial-analysis@claude-for-financial-services: the /comps and /dcf skills find an MCP data source and use it.

As a plain MCP server (Claude Code, Claude Desktop or any MCP client):

claude mcp add open-finance \
  -e SEC_USER_AGENT="Your Name you@example.com" \
  -- uvx --from git+https://github.com/nathanaelhub/open-finance-mcp open-finance-mcp

Responses are cached on disk (~/.cache/open-finance-mcp, or $OPEN_FINANCE_MCP_CACHE): company facts for a day, prices for 15 minutes. SEC requests are spaced to stay under the 10 requests/second fair-access limit.

How the numbers are built

XBRL data is messier than it looks. Each rule below exists because a real filer broke the simple version:

  • Concepts resolve per period, not per company. Tags change over time: Caterpillar's NetIncomeLoss stops in 2010 and continues as NetIncomeLossAvailableToCommonStockholdersBasic. Each metric has a priority list, the first concept with a value for that period wins, and the concept used is returned with the value.

  • Restatements win; labels come from the original 10-K. A 10-K repeats prior years as comparatives and may restate them. Values come from the most recent filing, and each value cites it. The fiscal-year label comes from the period's original 10-K, because a year-end date misleads for retailers (Home Depot's fiscal 2025 ends 1 Feb 2026).

  • LTM = FY + YTD − prior-year YTD, from one concept. JPMorgan tags annual revenue as Revenues and quarterly as RevenuesNetOfInterestExpense, so an LTM mixing concepts could combine different definitions. If no single concept covers all three periods, LTM is null. Labels give the exact end date, not "6M", because 52/53-week filers have 12- and 16-week quarters.

  • Total debt never double counts. DebtCurrent already includes commercial paper and the current portion of long-term debt; LongTermDebt already includes its current portion. The formula used is returned (e.g. ltd_noncurrent + ltd_current + commercial_paper), and matches Apple's FY2024 10-K to the dollar ($106.629B). Operating leases are excluded.

  • Missing is null, never zero. Caterpillar's 10-Qs tag debt only by segment, which the XBRL API omits, so LTM debt is null with a reason.

  • Share counts handle multiple classes. The cover-page count is summed across classes. Alphabet reports its cover page only per class, which the API drops, so the server falls back to the balance-sheet total and says the price is for one class.

  • Beta uses completed months. Yahoo's monthly series ends with the in-progress month; that partial "return" is dropped. Beta is OLS on the last 60 completed monthly returns against ^GSPC (a price index, no dividends).

  • Errors the model can act on. The MCP SDK hides unexpected exceptions behind "Error executing tool", so every anticipated failure is a readable tool error: an unknown ticker points to lookup_company, and a new registrant with no history (ExxonMobil's new holding company, CIK 2115436) is explained rather than returned as empty tables.

Earnings releases: the newest quarter

SEC's XBRL API can trail EDGAR by weeks (Coca-Cola's July 10-Q, above). The press release is furnished on results day as an exhibit to an 8-K tagged Item 2.02, so get_earnings_release finds the latest such 8-K, picks the EX-99 exhibit from the filing's index page (file names vary: a2026q2earningsreleaseex-9.htm, q2fy27pr.htm, …) and converts it to text. Financial tables come through as rows, with the $, ( and % cells that filings use for alignment joined back to their numbers:

EMEA | $3,240 | $3,176 | 2 | $1,309 | $1,325 | (1)
Consolidated | $13,380 | $12,535 | 7 | $4,672 | $4,280 | 9

Tested across Apple, Microsoft, NVIDIA, Alphabet, JPMorgan, Caterpillar, PepsiCo, Coca-Cola, Walmart and Tesla. Item 2.02 also covers non-earnings results announcements (Tesla files delivery numbers under it), so the tool tells the model to confirm what the document is rather than guessing.

Foreign filers (IFRS)

Foreign private issuers file 20-F/40-F reports under IFRS, which SEC's API publishes in a separate ifrs-full taxonomy, in each company's own currency. get_financials reads them with an IFRS concept map (profit attributable to owners of the parent, IFRS Borrowings as current + non-current debt) and returns taxonomy and currency with the figures: Toyota comes back in yen, SAP and Unilever in euros, Novo Nordisk in kroner. A filer with both taxonomies (Sony moved from US GAAP to IFRS) is read from the one with the newest year. get_fx_rate supplies the FRED rate to convert with: a period average for income-statement items, the rate on the balance-sheet date for cash and debt.

What it deliberately doesn't do:

  • No LTM. Interim results are furnished on 6-K without XBRL.

  • No market cap. A US listing is usually an ADR whose ratio of ordinary shares isn't in XBRL (one Toyota ADR is 10 shares), so price x shares can be off by that ratio. Price and share count are returned separately.

  • No cross-currency comps. get_comps refuses a non-USD reporter instead of dividing a dollar price by yen earnings.

  • No guessed concepts. Toyota tags capex only with its own extension, so capex and free cash flow are null, not estimated.

Limitations

  • IFRS filers have annual data only, and market cap must be computed by hand (see above). Captive-finance borrowings (carmakers) inflate net debt the way deposits do for banks.

  • The XBRL company-facts API excludes dimensional (segment/class) facts and can lag EDGAR by weeks. Both cases are flagged rather than hidden.

  • Prices come from Yahoo Finance's unofficial chart endpoint. They are labeled as unofficial so a model citing them says so; verify before publishing.

  • EBITDA is operating income + D&A as reported, not adjusted EBITDA.

Development

uv sync
uv run pytest            # 75 tests, offline: real filings trimmed into tests/fixtures
SEC_USER_AGENT="Name you@example.com" uv run python scripts/make_fixtures.py   # refresh fixtures

Tests run the tools through an in-process MCP client with HTTP mocked, and once over stdio as a subprocess, which is how Claude Code launches the server. CI covers Python 3.11–3.14.

Data: SEC EDGAR APIs, FRED. Not investment advice; outputs are drafts for review by a qualified person.

MIT licensed.

Available Tools

6 tools
get_compsB
Read-onlyIdempotent

Trading comparables: market cap, EV, LTM revenue/EBITDA/net income, EV/Revenue, EV/EBITDA, P/E.

EV = market cap + latest net debt. Multiples are null (with a reason) when an input is missing, negative, or not meaningful (financials).

Inputs: tickers (required).

get_filingsB
Read-onlyIdempotent

Recent SEC filings with document links. forms filters, e.g. ["10-K", "10-Q", "8-K"].

Inputs: ticker (required), forms, limit. Returns: result

get_financialsA
Read-onlyIdempotent

Standardized annual (and LTM) financials from SEC XBRL filings.

Income statement, cash flow and balance-sheet items plus derived EBITDA, free cash flow, total debt and net debt. Every value carries its XBRL concept and filing accession; filings maps accessions to URLs.

Inputs: ticker (required), years, include_ltm.

get_market_dataB
Read-onlyIdempotent

Share price, SEC cover-page shares outstanding, market cap and 5-year monthly beta.

Inputs: ticker (required).

get_treasury_yieldB
Read-onlyIdempotent

Latest US Treasury constant-maturity yield from FRED (percent). 10Y is the usual DCF risk-free rate.

Inputs: maturity.

lookup_companyB
Read-onlyIdempotent

Find SEC registrants by ticker or name. Returns ticker, name, CIK and exchange.

Inputs: query (required), limit. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.7/5.0

    Scored across 6 tools

    Disambiguation5/5

    Each tool serves a clearly distinct purpose: company lookup, market data, financial statements, SEC filings, Treasury yields, and trading comparables. Although get_market_data and get_comps both surface market cap, their scopes are well differentiated by the descriptions. No two tools appear to do the same thing.

    Naming Consistency4/5

    Most tools use a consistent get_<object> snake_case pattern, such as get_market_data, get_financials, and get_filings. lookup_company is the only deviation, using lookup_ instead of get_, but it still follows a readable verb_noun convention.

    Tool Count5/5

    Six tools is well-scoped for a focused open-finance data server. Each tool covers a distinct part of the financial-analysis workflow without redundancy. The count is neither bloated nor too thin.

    Completeness4/5

    The surface covers lookup, market data, financials, filings, risk-free rate, and comparables, which supports core fundamental workflows. Minor gaps remain, such as historical price series and quarterly financial statements, but agents can work around them with the available annual/LTM and current market data.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables agents to pull US stock financials, SEC filing timelines, ticker comparisons, screens, and fund profiles, with every value returned alongside the URL of the filing it came from. Answers are strictly filing-grounded and gaps come back as missing rather than estimated.
      68 npm
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables Claude to access free official financial and economic data from SEC EDGAR, U.S. Treasury, FRED, and the World Bank, including XBRL financial statements, key metrics, filings, insider trades, 13F portfolios, market prices, and macroeconomic series.
      20
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables free financial data analysis from SEC EDGAR filings, including company fundamentals, DCF valuation, comparable companies, industry benchmarks, and stock screening, with no API key required.
      81 npm
      1
      Apache 2.0