Canli Backtest MCP server

Canli Capital

Open quant research where every published number comes with the file that produced it, including the strategies that failed. Live at canlicapital.com.

Try it in a minute

# Ask your AI assistant whether a backtest is real (Claude Code shown; Cursor, VS Code and
# Claude Desktop setups are in mcp/README.md)
claude mcp add canli -- npx -y canli-validation-mcp

# Or call the free API: get a key, then deflate a Sharpe ratio for the number of trials you ran
curl -X POST https://canlicapital.com/api/v1/keys -H "Content-Type: application/json" -d '{"label":"readme"}'

# Or check the published record yourself (Python 3)
pip install cryptography
curl -sO https://canlicapital.com/glassbox/reproduce.py
for f in capacity_commitment founder_commitment; do curl -sO https://canlicapital.com/glassbox/$f.json; done
python3 reproduce.py --dir .

What is here

haircut Sharpe, luck-equivalent trials) as an MCP server and a free API, each result with a signed receipt anyone can recompute.

• Backtest validators (deflated Sharpe, CSCV overfitting probability, minimum track record,

(npx -y canli-fundamentals-mcp): what a company first reported, what was known on any date, and every later restatement, each value with its filing, computed locally from a hash-checked SEC snapshot.

• SEC fundamentals point in time as an MCP server

papers, killed candidates, trial counts and the live paper record.

• The research record as an MCP server (npx -y canli-research-mcp):

links to the filing it came from, with the year-on-year change and growth rate worked out.

• Company reference pages built from SEC XBRL filings: financial histories where every value

Null Zoo (overfitting corrections scored where the truth is known) and luck-equivalent trials.

• Research, including what failed: the kill log, the

questions with machine-checked answers from SEC filings (CC BY 4.0).

• Open data: FilingFacts, financial

github.com/arhancanli/alphac.

• ALPHAC, the three-strategy paper-trading book whose record is published daily; its engine is

Help check it: open review tasks take 15 to 60 minutes and are credited on /review. To build with it, see CONTRIBUTING.md. If the project is useful to you, a star helps other people find it.

Shortened. The full README is on GitHub.