Canli Quant

canli-quant-mcp

236 quant finance tools in one MCP server: performance and risk, options and exotics, fixed income, portfolio construction, econometrics, technical indicators, execution, sizing, crypto and FX, valuation, strategy backtests, a library of 399 strategy sleeves, data checks, and data annotation and ML labeling. Every tool computes on the data you send, and every tool's results are checked in the test suite against an independent implementation (QuantLib, statsmodels, arch, TA-Lib, numpy-financial, scikit-learn, empyrical, scipy, pandas, dcor): 465 reference cases, plus planted-defect tests for the data checks.

Local stdio, no key, no account, no network calls: your data never leaves your machine (see Privacy, enforced by tests).

Computed NaN and Infinity values are refused with a tool error naming the affected fields. They do not become unexplained nulls or receive a success receipt. Check the input scale and the calculation's domain. Ordinary finite answers and null warm-up positions are preserved.

[canli-quant-mcp on canlicapital.com](https://canlicapital.com/mcp-servers/quant) · All Canli Capital MCP servers. Free, MIT-licensed. For every server behind one install, use canli-mcp.

Three tools in context, 235 behind them

Listing 238 tools directly costs 83,944 tokens of context (o200k) on every request. By default this server lists three, 680 tokens:

find_tool ranks with BM25 over each tool's name, title, keywords and description, with stemming and finance abbreviations (cvar, npv, irr, ytm...). On 110 test requests it returns the right tool first 94% of the time and in the top three 99% (test/search.test.mjs), so a model rarely needs a second search. run_tool takes digits (3-10 significant figures; default 10) to cut output tokens when full precision is not needed: 9-19% fewer tokens at 4-6 digits on the results we measured.

Measured on an Apple-silicon laptop (Node 24): connect and initialize 140 ms; a typical tool call 0.26 ms median over stdio; find_tool 1 ms; a full 399-sleeve tournament on 1,200 days of six assets about 2 s.

The tool list is byte-identical across launches, so providers can cache it. To list toolsets directly instead, set CANLI_TOOLSETS to a comma-separated list (e.g. performance,options, 8,053 tokens for performance alone) or all.

Quick start

npx -y canli-quant-mcp

Claude Code:

claude mcp add canli-quant -- npx -y canli-quant-mcp

Claude Desktop, Cursor or any MCP client:

{ "mcpServers": { "canli-quant": { "command": "npx", "args": ["-y", "canli-quant-mcp"] } } }

Then ask, for example:

• *"Here are my strategy's daily returns. I tried 40 variants. Does the edge survive deflation?"*
• *"Price a 6-month down-and-out call, strike 100, barrier 90, vol 25%, and compare it with the vanilla."*
• *"Is this spread mean-reverting? Run ADF, KPSS and the variance ratio, then backtest it with costs."*
• *"Check this option chain for arbitrage before I fit a surface to it."*
• *"Which of the 399 sleeves hold up on these ETF prices after correcting for trying them all?"*
• *"Is this market in a calm or a turbulent regime right now, and how long do regimes last?"*

Toolsets

Workflows (prompts)

Prompts chain the tools for common jobs: analyze_strategy, audit_backtest, value_company, price_option, test_mean_reversion, build_portfolio, analyze_trade_log, sleeve_tournament, build_sleeve_book, label_dataset, check_dataset.

Resources: canli-quant://catalog (every tool), canli-quant://methods (what each toolset is checked against), canli-quant://tools/{name} (one tool's schema), canli-quant://sleeves (the sleeve library), canli-quant://sleeves/{id} (one sleeve's spec) and canli-quant://privacy.

Sleeve library

399 strategy sleeves in 18 families, each a published idea with its usual parameters swept on a plain grid: moving-average crossovers, time-series momentum, multi-horizon trend, channel breakouts, MACD, z-score and RSI reversion, volatility targeting, cross-sectional momentum, dual momentum, risk parity, equal weight, minimum variance, low volatility, short-term reversal, trend-filtered allocation, 52-week-high momentum and pairs. Every sleeve has a fixed rule, a stated warm-up, references, known risks and a SHA-256 of its spec. None was picked because it backtested well: they are definitions, not track records.

Running many strategies and keeping the best is how backtests overfit, so the tools that run many sleeves at once correct for it:

The tests check that no sleeve looks ahead (rewriting future prices never changes an earlier target), that none trades before its warm-up, and that the tournament's statistics match arch's Reality Check and an independent re-studentized SPA and StepM on the same bootstrap draws, scipy's clustering and an independent numpy CSCV. run_sleeve and combine_sleeves return target_weights by symbol, which a paper broker such as canli-paper-trading-mcp can rebalance to.

Data annotation and ML labels

The labeling toolset annotates your own data and turns it into training labels without leakage:

Each label says which future periods it used, so purging and embargoing are exact. The methods follow Lopez de Prado's *Advances in Financial Machine Learning*, checked against pandas ports of its code.

Privacy

Your data stays on your machine. The server is a local stdio process with no network code, writes nothing to disk, logs nothing, starts no processes, and keeps arguments only for the duration of a call. test/privacy.test.mjs holds the code to that with a static scan of every source file and by running the server under Node's permission model, with reads limited to its own package and writes, child processes and workers denied. For the same lockdown in use:

npm i -g canli-quant-mcp

node --permission --allow-fs-read="$(npm root -g)/canli-quant-mcp" "$(npm root -g)/canli-quant-mcp/src/server.mjs"

canli-quant://privacy states the guarantees from inside the server. Calculation receipts carry hashes, never the data.

Calculation receipts

run_tool with receipt: true returns the SHA-256 of the tool name and arguments, the SHA-256 of the result, and the server version. Anyone can run the same tool with the same arguments on that version and compare the output hash. Results are deterministic: no clock, no network, and the only randomness is the sleeve bootstrap, drawn from a stated seed.

Conventions

• Returns are simple fractions (0.01 = 1%), oldest first. Tools refuse returns of -100% or below and series that look like percents.
• Rates are annual; options use continuous compounding; bond yields compound at the coupon frequency.
• Results are rounded to 10 significant figures.
• Strategy backtests decide weights at the close of period t and earn period t+1's return; weights drift between rebalances and costs are charged on turnover.

All tools

performance : Performance and risk (23)

| max_drawdown | Find the maximum drawdown and the worst draw