00Canli Capital
Finance runs on numbers. Left alone, AI models make them up.
Canli Capital is an open-source lab that connects AI agents to financial data and quant tools they can check. Its flagship MCP server, canli-mcp, reads SEC filings, market data and economic statistics at the source and puts 388 finance tools behind five, with interactive charts in the chat. It is free, MIT-licensed and needs no API key or account.
01The problem
Ask a model for a company's numbers, and it guesses.
We asked AI models 150 questions about what companies reported to the SEC, with no tools. GPT-5.4 mini gave 84 numbers. 96% of them were wrong.
The wrong numbers read exactly like right ones. Nothing in the answer tells you which is which.
- 84numbers given
- 96%of them wrong
- 150questions
02The tools
The open tools hand the model a haystack.
OpenBB's MCP server lists 1,056 tools. Sent in full, they cost 432,001 tokens on every request, before the model reads the question.
And in our benchmark, every other server missed at least one kind of question entirely: SEC filings, rates and macro, prices or analytics.
- 1,056tools in one list
- 432,001tokens per request
- 4 of 4left a kind of question at zero
03The solution
One server between the model and the market.
canli-mcp shows the model 5 tools and keeps 388 behind them. The model searches for what it needs, reads that tool's arguments and runs it. The data comes from where it was first published: SEC filings, the Treasury's yield curve and FRED.
find_toolsearches every tool by what you ask fordescribe_toolreads one tool's argumentsrun_toolruns any tool, or up to 25 in one round trip, and draws the result as a viewhistorya company's reported figure over time, as first reported or as known on a datefinancial_statementsa company's statements exactly as filed
Give a model the filings and the numbers change. With a Canli MCP server reading them, gpt-5-mini answered 72% of the FilingFacts questions correctly, against 19% on its own.
04The receipt
An answer you can trace to the page it came from.
Ask for NVIDIA's head count. canli-mcp opens the latest 10-K, finds the line and returns it with the filing's accession number, its link and a SHA-256 hash of each document it read.
- asked
- NVDA · 10-K · find "employees"
- filed
- accession
- 0001045810-26-000021
- found
- "As of the end of fiscal year 2026, we had approximately 42,000 employees in 38 countries"
- evidence
- 3 documents, SHA-256 each · f192181c12dad789…
- time
- 764 ms
05The flagship
canli-mcp
Every Canli Capital server in one process. 5 tools in front, 388 behind them, 2,524 tokens of context, ready in 0.16 seconds.
- 235Quant financeperformance and risk, options, fixed income, portfolios, econometrics, indicators and a library of 399 strategy sleeves
- 17Backtest validationdeflated Sharpe ratio, overfitting probability, data-snooping tests, leakage checks and placebo tests
- 86MarketsSEC filings and statements as filed, foreign filers too, insider trades, 13F and 5% holders, fund portfolios, options chains and volatility surfaces, the Fed, Treasury yields and auctions, FRED, World Bank and OECD data, central banks abroad, prediction markets, prices and a whole-market overview, analysts' targets and technical indicators, a figure traced down its filing, what a token's contract can do, every annual report searched at once, management's record against its own guidance, what followed every past market fall, and ship traffic from satellite tracking
- 7SEC fundamentalscompany financials as first reported, with later restatements flagged
- 6Research recordCanli Capital's open research record
- 5Paper tradingAlpaca paper orders behind pre-trade checks, each check signed
- 32Factor backtestsfactor backtests that refuse lookahead and fill only what the market can take, crisis replays, purged cross-validation, a breadth lab that counts its independent bets, parameter sweeps on every core, data annotation, a signed ledger of filings, a paper league, a paper exchange where AI agents trade in order books, a count of how many independent views a group of models holds, and a search for what diversifies a book
Ask for a tool in plain words and the right one comes first 46 times in 48. Pass a file instead of pasting data, and 1,000 prices cost 23 tokens instead of 4,893.
06The gap
Same model. Same questions. Only the server changed.
On the 14 questions every server finished, canli-mcp answered 79% correctly. The best of the others, EdgarTools + Yahoo Finance MCP, answered 54%.
Interim result of 9 October 2026: gpt-5.4-mini, 2 runs per question, no API keys for any server. We wrote the questions, in areas our servers were built for, and what the others do well beyond them, such as news and options, is not scored.
07In the open
We hold our own strategies to the same standard.
ALPHAC, Canli Capital's trading engine, runs its strategies on paper and publishes every position, decision and broker reconciliation, daily. Its backtests face the same tests canli-mcp ships, such as the deflated Sharpe ratio, and the trials that failed stay on the record.
08Company data
US public companies, each figure traced to its filing.
Look up what a company reported, with every figure linked to the SEC filing it came from. An agent can ask for a value as it stood on a past date, with later restatements flagged.
09The builder
Built by Arhan Canli.
Founder and quantitative researcher, in Dubai, building since July 2024. Canli Capital began as an open lab for testing trading strategies honestly and grew into the tools AI agents need to work the same way.
The site keeps the tools' rule. Every figure on it traces to a published file, and corrections stay visible.
10The vision
Picture the market at dawn, every agent working from checked numbers.
Agents will research, test and trade at machine speed. Canli Capital is building the research platform they work on, in the open, in five branches that feed one destination.
- MCP servers and APIPrimary-source data and quant tools for agents, for few tokens.Live
- Data and annotationDatasets to train and test models, starting with FilingFacts. Expert labels come next.Started
- The sleeve libraryWhere developers build, train and test millions of strategy sleeves, with honest statistics on each.Started
- ALPHACOne book from hundreds of thousands of weakly correlated sleeves. Correlation sets the ceiling.Paper only
- The breakthroughs labNew statistics that tell a real edge from luck, published with their code.Started
- The flagship fundsFrom the US, trading the best of all five. Paper only until a licensed entity exists.Paper only
11Start
Give your agent the market.
canli-mcp is free and open source under the MIT license. Add it to Claude Code, Cursor or any MCP client in one line.
claude mcp add canli -- npx -y canli-mcp{ "mcpServers": { "canli": { "command": "npx", "args": ["-y", "canli-mcp"] } } }Side by side
canli-mcp, OpenBB, EdgarTools and Yahoo Finance MCP
Each of these open-source servers gives an AI model finance data in its own way. This is what each one offers, checked against its own tool list: canli-mcp 0.16.0 and the other servers in the versions we benchmarked. How each one scored is in the benchmark above, which tested canli-mcp 0.1.0.
| Question | Canli Capital | OpenBB | EdgarTools | Yahoo Finance MCP |
|---|---|---|---|---|
| The version compared | canli-mcp 0.16.0 | openbb-mcp-server 2.0.1 with openbb 5.0.0 | edgartools 5.61.1 | Alex2Yang97/yahoo-finance-mcp at commit b9c1765 |
| Tools sent to the model on every request | 5, in front of 388 | 1,056, or 14 in its tool-discovery mode | 13 | 9 |
| Finance question types it has a keyless tool for, of 53 | 51 | 39 | 11 | 13 |
| SEC filings, Form 4 insider trades and 13F holdings | Yes, read from EDGAR | Yes, through its SEC tools | Yes | Yahoo's holder and insider summaries, not the filings |
| Treasury yields and FRED series | Yes, with no API key | Yes; its FRED tools need a FRED API key | No | No |
| Quant analytics and backtest validation | Yes: 235 quant tools and 17 validation tools, run on your machine | None in the version tested | No | No |
| News, options chains and analyst ratings | Options chains and volatility surfaces from Cboe's delayed data and news from the GDELT open index, no key; no analyst ratings | News and options tools, some needing provider keys | No | Yes |
| License | MIT | Apache-2.0 | MIT | MIT |
Checked on against canli-mcp source, the coverage map (53 finance questions, each server's own tool list), OpenBB MCP server, EdgarTools, Yahoo Finance MCP, the head-to-head benchmark and its tool counts. Each project changes over time; its own pages are the authority.
Questions
Plain answers.
What is Canli Capital?
Canli Capital is an open-source lab that connects AI agents to financial data and quant tools they can check. It publishes MCP servers, company financials traced to SEC filings, open benchmarks and its own paper-trading record, failures included. See what stays open.
What is canli-mcp?
canli-mcp is one MCP server that loads every Canli Capital server in one process. The model sees five tools, find_tool, describe_tool, run_tool and the two it asks for most, and behind them sit 388 tools for quant finance, backtest validation, SEC filings and fundamentals, ownership, funds, options, the Fed, economic data, prices and paper trading. See every server it combines.
How do I add canli-mcp to Claude, Cursor or another AI assistant?
In Claude Code, run claude mcp add canli -- npx -y canli-mcp. In Cursor, Claude Desktop or any other MCP client, add a server named canli whose command is npx with the arguments -y and canli-mcp. It needs Node.js 20.10 or later and no account. Follow the setup guide.
How does canli-mcp compare with OpenBB, EdgarTools and Yahoo Finance MCP?
With the same model and questions, canli-mcp answered 79% of the questions every server finished correctly, against 54% for the best of the others, in the version tested, which sent the model 859 tokens of tools on every request; canli-mcp 0.12 sends 2,524 with more tools and charts. We wrote the questions, so read the caveats too. See every run and the caveats.
Where does the data come from?
From the original publishers: companies' filings with the SEC, the US Treasury's daily yield curves, FRED for economic series and Yahoo Finance chart data for prices. Answers drawn from data name their source, and answers from a filing carry its accession number. Browse the company pages.
How do I check whether my backtest is overfit?
Send the returns your backtest produced to the validators. The deflated Sharpe ratio corrects for how many variants you tried, the probability of backtest overfitting compares those variants with each other, and the minimum track record length says how much history the Sharpe ratio needs. Read the step-by-step guide.
Is it free?
Yes. The MCP servers and the code are free and MIT-licensed, the published data is CC BY 4.0, and the hosted API has fair-use limits. No paid plan is offered. Start with the developer quickstart.
Does my data leave my computer?
The quant and validation tools run on your machine and send nothing anywhere. The data tools fetch public data from the SEC, the Treasury, FRED, Yahoo Finance or canlicapital.com, and send only the company, series or symbol you asked for. See the setup and local mode.
Are your strategies trading real capital?
No. ALPHAC, Canli Capital's trading engine, runs its strategies on paper, and the record says so on every page. Paper results do not establish what real trading would return. Inspect the record and its basis.
Can I invest or copy the trades?
Canli Capital is a research project, not a fund or investment adviser. This site does not offer managed capital or copy-trading. Read the project's boundaries.
Release notes
Hear when it ships.
Everything here is free and needs no account. Leave an email only if you want a note when a server or a benchmark is released.