Skip to content

Benchmark, interim result of

Finance MCP servers, head to head

On the 14 questions every server finished, canli-mcp answered 79% correctly (22 of 28 runs). The best of the other servers, EdgarTools + Yahoo Finance MCP, answered 54% (15 of 28).

The setup

Same model. Same questions. Only the server changed.

Every server got the same model (gpt-5.4-mini), system prompt, 12-turn limit, 30,000-character cap on each tool result and scoring. No server had an API key. Each question ran 2 times per server, and the answers were computed from primary sources (SEC EDGAR, the US Treasury, FRED and Yahoo Finance chart data) separately from every server compared.

Accuracy

Correct on the questions every server finished

  1. canli-mcp79%22 of 28answered every kind of question
  2. EdgarTools + Yahoo Finance MCP54%15 of 28never answered: analytics
  3. Yahoo Finance MCP50%14 of 28never answered: analytics
  4. OpenBB MCP39%11 of 28never answered: SEC filings, rates and macro · not run: prices, analytics
  5. EdgarTools MCP39%11 of 28never answered: SEC filings, rates and macro, prices, analytics
Correct runs on the 14 shared questions, 2 runs each
ServerCorrectRuns
canli-mcp canli-mcp 0.1.0 as it stood before the benchmark, MIT79%22 of 28
OpenBB MCP openbb-mcp-server 2.0.1 with openbb 5.0.0, in its tool-discovery mode, Apache-2.039%11 of 28
EdgarTools MCP edgartools 5.61.1, its MCP server, MIT39%11 of 28
Yahoo Finance MCP Alex2Yang97/yahoo-finance-mcp at commit b9c1765, MIT50%14 of 28
EdgarTools + Yahoo Finance MCP both servers connected at once, MIT54%15 of 28

Coverage

By kind of question, over every finished run

Red marks a kind of question a server never answered correctly.

Correct runs by category, median input tokens and median seconds per run
ServerSEC filingsFundamentalsInsider trades13F holdingsRates and macroPricesAnalyticsInput tokensSeconds
canli-mcp5/63/66/64/68/86/66/1011,7825.33
OpenBB MCP0/66/63/62/60/4not runnot run107,112.524.44
EdgarTools MCP0/65/62/64/60/80/60/318,5589.86
Yahoo Finance MCP2/66/62/62/62/84/60/1010,3353.51
EdgarTools + Yahoo Finance MCP1/66/62/64/62/84/60/321,18610.37

Cost

What each server costs before the model asks anything

Each server is started with the benchmark's command; its instructions and whole tool list are counted in the shape an OpenAI-style client sends tools. No model is called. Tokens are o200k_base, measured 2026-10-11.

Tools sent to the model, tools reachable, tokens on every request, seconds to start and list
ServerTools sentTools reachableTokensSeconds
canli-mcp33068590.08
OpenBB MCP, every tool1,0561,056432,00112.03
OpenBB MCP, tool discovery (as benchmarked)141,0562,17110.32
EdgarTools MCP13133,8011.17
Yahoo Finance MCP992,2050.91

Before quoting it

Read this before quoting it

Every question

Every question, every server

Correct runs per question for every server, including the two tuned canli-mcp columns: canli-mcp after first fixes and canli-mcp after all fixes.

Correct runs per question and server
Questioncanli-mcpcanli-mcp after first fixescanli-mcp after all fixesOpenBB MCPEdgarTools MCPYahoo Finance MCPEdgarTools + Yahoo Finance MCP
According to NVIDIA's most recent annual report on Form 10-K, approximately how many employees did NVIDIA have at the end of the fiscal year? Answer with the number of employees. Answer: 42,0001/22/22/20/20/22/20/2
Tesla filed a press release on production and deliveries for the third quarter of 2026 with the SEC (an 8-K exhibit). How many vehicles did it say Tesla delivered in Q3 2026? Answer with the number of vehicles. Answer: 486,0002/22/22/20/20/20/20/2
On the cover page of Apple's annual report on Form 10-K for fiscal year 2025, how many shares of common stock were outstanding? Answer with the number of shares. Answer: 14,776,353,0002/22/22/20/20/20/21/2
What were Apple's total net sales (revenue) for fiscal year 2025, which ended September 27, 2025? Answer in US dollars as a plain number. Answer: 416,161,000,0001/22/22/22/22/22/22/2
What was Microsoft's net income for its fiscal year ended June 30, 2025? Answer in US dollars as a plain number. Answer: 101,832,000,0001/21/22/22/21/22/22/2
What was NVIDIA's diluted earnings per share for its fiscal year 2026 (ended January 25, 2026)? Answer in US dollars per share. Answer: 4.91/22/22/22/22/22/22/2
According to SEC Form 4 filings, how many shares of Tesla common stock did Tesla's Chief Financial Officer sell in open-market sales on September 8, 2026? Answer with the number of shares. Answer: 2,605.752/21/21/22/22/22/22/2
Add up all open-market sales (Form 4 transaction code S) of NVIDIA common stock by NVIDIA insiders reported in Form 4 filings filed between July 1 and September 30, 2026. How many shares were sold in total? Answer with the number of shares. Answer: 4,332,1182/22/22/21/20/20/20/2
How many Form 4 filings were filed with the SEC for Apple Inc. as the issuer during September 2026 (filing dates September 1-30, 2026, original Form 4s only, not amendments)? Answer with the count. Answer: 112/22/22/20/20/20/20/2
According to Berkshire Hathaway's 13F filing for the quarter ended June 30, 2026, how many shares of Apple did it hold? Answer with the number of shares. Answer: 227,917,8080/22/22/22/22/22/22/2
What is the total market value of all holdings reported in Berkshire Hathaway's 13F filing for the quarter ended June 30, 2026? Answer in US dollars as a plain number. Answer: 299,253,556,2462/22/22/20/22/20/22/2
How many entries (rows in the information table) does Bridgewater Associates' 13F filing for the quarter ended June 30, 2026 report? Answer with the count. Answer: 9972/21/22/20/20/20/20/2
What was the 10-year US Treasury par yield (Treasury daily par yield curve rate) on October 8, 2026? Answer in percent, e.g. 4.25. Answer: 5.222/22/21/20/20/22/22/2
Using the US Treasury daily par yield curve, what was the 10-year yield minus the 2-year yield on October 8, 2026? Answer in percentage points, e.g. 0.35. Answer: 0.472/20/22/20/20/20/20/2
What was the year-over-year change in the US Consumer Price Index for All Urban Consumers (CPI-U, all items, seasonally adjusted) for August 2026? Answer in percent, e.g. 2.9. Answer: 3.3532/22/21/2not run0/20/20/2
What was the US unemployment rate for August 2026 (seasonally adjusted)? Answer in percent. Answer: 4.12/22/22/2not run0/20/20/2
What was Apple's (AAPL) closing stock price on October 8, 2026? Answer in US dollars. Answer: 340.422/22/22/2not run0/22/22/2
What was Microsoft's (MSFT) total return including dividends from the close on December 31, 2025 to the close on September 30, 2026? Answer in percent, e.g. 12.3. Answer: 6.72722/22/21/2not run0/21/20/2
What was the maximum drawdown of SPY during calendar year 2025, measured on daily dividend-adjusted closing prices (peak to trough within 2025)? Answer as a fraction, e.g. 0.12 for 12%. Answer: -0.1875522/22/22/2not run0/21/22/2
What was NVIDIA's (NVDA) annualized volatility in 2025, computed as the standard deviation of daily simple returns of adjusted closing prices times the square root of 252 (returns from the first trading day of 2025 through December 31, 2025, starting from the December 31, 2024 close)? Answer as a fraction, e.g. 0.35. Answer: 0.496182/22/22/2not run0/20/20/2
What was the correlation between the daily simple returns of AAPL and MSFT in calendar year 2025, using adjusted closing prices? Answer as a number between -1 and 1. Answer: 0.4869140/21/22/2not run0/10/20/1
What was Tesla's (TSLA) beta to SPY in calendar year 2025, from daily simple returns of adjusted closing prices (covariance over SPY variance)? Answer as a number. Answer: 2.2343810/20/22/2not runnot run0/2not run
What was SPY's annualized Sharpe ratio in calendar year 2025, with a 0% risk-free rate, from daily simple returns of adjusted closing prices (mean over standard deviation times the square root of 252)? Answer as a number. Answer: 0.94012/22/22/2not runnot run0/2not run
Price a European call option with Black-Scholes: spot 100, strike 105, 6 months to expiry, risk-free rate 4% (continuous), dividend yield 1% (continuous), volatility 22%. Answer in dollars. Answer: 4.7122442/22/22/2not runnot run0/2not run

Reproduce

Run it yourself

The harness, the questions, the code that computes each answer and every run with its tool calls are in mcp-canli/bench/rivals at commit 8c106a6f. All figures on this page are in finance-mcp-servers.json.