The vision
Picture the market at dawn, every agent working from checked numbers
AI agents will research, test and trade at machine speed. Canli Capital is building what they stand on, in the open, in four parts: the context they read, the tests that keep them honest, the data they learn from and the execution they act through.
- 306finance tools behind one MCP server
- 8servers of the family on npm
- 6,588npm downloads in the last 30 days
- 79%correct in the open benchmark; the best of the others, 54%
- 72%of FilingFacts questions right for gpt-5-mini with a Canli MCP server, against 19% on its own
Four parts
What it stands on
Financial software was built for people at terminals. More and more of the work is now done by AI agents. Everything here is open source and published as it is built, including the tests that fail; each part says what exists today and what comes next, and a goal stays a goal until it is built and measured.
01
Context
LiveMCP servers that give AI agents primary-source financial data and quant tools, for as few tokens as possible.
Today
- canli-mcp puts 306 tools behind 3: SEC filings, fundamentals, Treasury and FRED data, prices, quant analytics, backtest validation and paper trading, in 859 tokens of context.
- 8 servers of the family are on npm, three of them also hosted at canlicapital.com/mcp.
- In the open benchmark, canli-mcp answered 79% of the questions every server finished; the best of the other open servers, 54%.
Next
canli-mcp on npm and the MCP Registry, a held-out benchmark set the servers were not tuned on, and more servers in the family.
02
Testing
LiveAn engine that counts every trial, so a lucky backtest is called luck instead of a strategy.
Today
- ALPHAC, the open-source engine, keeps its paper-traded record, its trial accounting and its retracted figures in public.
- The validators (deflated Sharpe, probability of backtest overfitting, minimum track record, the haircut Sharpe and White's Reality Check) are checked against their papers. They measure overfitting risk; they cannot remove it.
Next
More asset classes and more economically distinct strategies, each admitted only on forward evidence.
03
Data
StartedFinancial datasets to train and test AI models, where every answer is computed from public filings and checked by qualified people.
Today
- FilingFacts asks AI models questions about what companies reported to the SEC. On their own, gpt-5-mini answered 19% correctly; with a Canli MCP server reading the filings, 72%.
- Every answer is recomputed by an independent checker and cites its filing. No item is called human-verified until two people have checked it.
Next
The first double-labelled expert gold set, and benchmark results for more models.
04
Execution
Paper onlyStrategies that pass the tests, run by agents through real brokers, with every order journaled and checkable.
Today
- A public paper-traded record, not funded, with every position, decision and broker reconciliation published.
- canli-paper-trading-mcp places Alpaca paper orders behind pre-trade checks and a kill switch.
Next
Real capital only after licensing, a legal entity and a forward record that justifies it. Nothing here is investment advice or an offer.
Follow and use it
Build on it
Install canli-mcpDeveloper guideUsageSource on GitHubThe founder