Canli Capital

canli-mcp

Every Canli Capital MCP server in one: 306 finance tools for Claude, Cursor or any MCP client.

• Quant finance: 235 tools covering performance and risk, options and exotics, fixed income, portfolios, econometrics, indicators, and a library of 399 strategy sleeves with multiple-testing corrections.
• Backtest validation: deflated Sharpe, overfitting probability, data-snooping tests, leakage checks and placebo tests, run on your machine.
• Point-in-time SEC fundamentals.
• Markets from the sources, no key: screens of every US-listed company, company reports, event studies, SEC filings and their sections, insider trades, 13F and 5% holders, fund and ETF portfolios, index members, short interest, off-exchange volume, fails to deliver, futures positioning, executive pay, revenue by segment, options chains and volatility surfaces, the VIX futures curve, Fed decisions and projections, the economic calendar, Treasury yields and auctions, the federal debt, FRED, World Bank and OECD data, and prices for stocks, currencies, futures, indices and crypto.
• Canli Capital's open research record.
• Factor backtests without lookahead.
• Alpaca paper trading behind pre-trade checks.

Each pack is a published, separately tested server. This package loads them into one process and puts three tools in front of them.

npx -y canli-mcp

Claude Code:

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

Claude Desktop, Cursor or any MCP client:

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

Why one server is cheaper and faster

Measured on Node 24 on 2026-10-11 with bench/context.py (o200k tokens of the tool objects a client receives, plus the server instructions; start-up is the median time from launch to the tool list):

• Three tools in context. The model searches with find_tool, reads one schema with describe_tool, and calls anything with run_tool. On 48 everyday requests written after the last change to the search and never tuned on, the right tool comes first 46 times and is in the top three 47 times (0.2.1: 25 and 29). test/search.test.mjs holds these and 135 more.
• Nothing loads until it is used. Start-up reads a prebuilt index of all 306 tools; a pack's code loads the first time one of its tools runs.
• Data by file, not by pasting. Any argument can be {"$file": "prices.csv", "column": "close"} (or "columns": ["SPY", "TLT"] or "all"). For 1,000 prices that is 23 tokens instead of 4,893, and the model cannot drop a number while copying.
• Several calls in one round trip. run_tool takes calls (up to 25). A call can use an earlier call's output with {"$result": 0, "path": "rows", "pick": [0, 1]}, so intermediate data never passes through the model, and return: "last" sends back only the final answer. Labeling 1,000 prices and then weighting the labels took 72 tokens of tool input this way, against 31,917 to read the labels and type them into the next call.
• Shorter numbers when they are enough. digits: 5 rounds every result.

Workflows as prompts

Clients that show MCP prompts (Claude Desktop and Claude Code list them as slash commands) offer five workflows. Each one runs a single batched run_tool call and says what to write from the results:

A prompt is offered only when every tool it names is in an enabled pack (CANLI_PACKS, CANLI_OFFLINE), and test/prompts.test.mjs checks each call against its tool's own schema. Prompts add nothing to a request's context until one is chosen.

An agent, end to end

The same model (gpt-5.4-mini) answered seven tasks twice each:

• the Sharpe ratio, drawdown and value at risk of a pasted price file;
• a deflated Sharpe from summary numbers;
• a Black-Scholes price;
• a 399-sleeve tournament on three assets;
• Apple's first-reported 2019 EPS.

Answers were graded against the tools' own values (bench/agent-eval.mjs; results in bench/agent-eval-2026-10-08.json):

The fixes came from reading the runs:

• The model narrowed find_tool to one pack and missed the tool that took its inputs, so the pack filter is gone.
• It passed made-up two-point series to a tool that wanted returns, so the quant deflated and probabilistic Sharpe tools now take summary numbers and refuse too little data.
• It put select and return inside a tool's arguments, so run_tool moves them out.
• It asked select for fields that do not exist, so the reply now lists the fields that do.

This is one small model with two runs per task, so read the differences as large or small, not as exact rates.

Privacy

Everything runs on your machine. canli://privacy states, per pack, what leaves it:

CANLI_OFFLINE=1 enables only the packs with no network, and no other pack's code is even loaded. test/privacy.test.mjs runs the server that way with fetch and every network module made to throw, and the quant and validation tools still answer. The server itself has no network, write or logging code (also tested). Opt-in receipts (receipt: true) carry hashes, never data.

Configuration

Order-sending tools never run inside a batch: preview first, then send on its own.

Packs and what each is checked against

Two tools differ from their standalone servers. Validation's stress_test is strategy_stress_test here, because the quant pack has a portfolio stress_test. Validation's get_key and service_status are left out, because validation runs locally here. canli://packs lists both.

Development

npm ci && npm testnpm run build-index        # after updating a pack; a test fails if the index and the packs differnode bench/robustness.mjs  # every tool, offline, with hostile arguments: crashes and leaks reported

MIT licence.