Canli Paper Trading

canli-paper-trading-mcp

Lets an AI agent trade your Alpaca paper account without being able to skip the checks. Every order is previewed against your own limits, the market clock and a kill switch; only orders that pass get a five-minute confirmation token, and only a token can send them. Paper only: the package contains no live trading host, and live keys are refused.

Local stdio. Your keys stay in your environment; nothing goes anywhere except Alpaca's paper API.

Tools

The checks

The pre-trade checks are canlicapital's shared core (js/pretrade-core.js, mirrored byte for byte and tested in the repository). Per order: kill switch, allowed symbols and order types, market open or closed, price collar, order and daily notional caps, position, gross and net caps against your whole book, reduce-only integrity, and an estimated cost (half spread; commission and impact when you supply fees and volume). A rejection lists every reason; nothing is resized.

Your limits live in CANLI_HOME/limits.json (default ~/.canli), for example:

{ "max_order_notional": 10000, "max_position_frac": 0.2, "max_gross": 1.0, "allowed_types": ["limit"] }

A call can tighten these for one preview, never loosen them. While a file named KILL exists in that directory, every preview fails and every send is refused.

Tokens

• Bound to the exact orders previewed: change a quantity and the token no longer verifies.
• Single use and valid for five minutes, so a send never acts on stale checks.
• Signed with a key that lives only in the running server; a restarted server rejects old tokens.
• Each sent order carries a client_order_id derived from its token. Alpaca refuses a repeated id, so a retried send cannot fill twice.

Order log

Every send and cancel is appended to CANLI_HOME/paper-orders.jsonl, each line holding the hash of the one before. paper_account reports whether the chain still verifies; an edited line breaks it.

Signed attestations and private balances

Every preview, rebalance and send returns an attestation: which orders were checked (their hash), against which limits (their hash), the verdict, the kill switch, the time, and commitments to the account's equity and positions, signed with an Ed25519 key made on your machine the first time (CANLI_HOME/attest-ed25519.pem, mode 0600; it never leaves the machine). Give attest-ed25519.pub.pem to an auditor or counterparty; verify_attestation checks a statement was signed by that key and not altered. On Node 24.7 or later each statement is also signed with an ML-DSA-65 key (FIPS 204, a post-quantum signature; attest-ml-dsa-65.pem), over the same statement naming both keys: give both public keys, and both signatures must hold, so a statement stays unforgeable even if one scheme is broken.

Set CANLI_PAPER_PRIVATE=1 and the model never sees the account's figures: paper_account returns commitments (SHA-256 of a random salt and the value) instead of equity, cash, buying power and positions; rebalance_to_weights sizes orders without showing them; numbers in check messages are blanked. The salts are written only to CANLI_HOME/commitments.jsonl, so you can later open a commitment to someone (show its value and salt) and verify_attestation confirms it.

This is a signature over commitments, not a zero-knowledge proof: whoever you open a value to learns it, and the checks are this server's code. Brokers accept API keys, not proofs; the keys stay in the environment and are never shown to the model in either mode.

Quick start

1. Create an Alpaca account and open the paper trading dashboard; generate paper API keys (the key id starts with PK).
2. Add the server with your keys in its environment:

claude mcp add canli-paper -e ALPACA_PAPER_KEY_ID=PK... -e ALPACA_PAPER_SECRET_KEY=... -- npx -y canli-paper-trading-mcp

Then ask: "Rebalance my paper account to 60% SPY and 40% TLT, show me the checks, and send it if everything passes."

Pairs with canli-quant-mcp: backtest a strategy there, then move the paper account to its target weights here.

What this does not do

It does not trade real money, and there is no setting that makes it. It does not decide what to trade. Paper fills are Alpaca's simulation and are kinder than real markets (no queue position, no market impact); a strategy that works on paper has not yet shown it works live.

MIT licence.