Underwrite

The options agent that has to earn its own position size.

Every autonomous trading agent in this hackathon claims it manages risk. Underwrite measures whether its own claims are true, and will not size up until they are.

• Strategist — a Gemini agent (Google ADK) whose only market access is Alpaca's MCP server, filtered to read-only tools. It proposes one defined-risk options structure per cycle with a stated probability of profit, or declines.
• Second opinion — a different model (Claude Sonnet 4.5 via OpenRouter, or an open-source model on Featherless) reviews the specific legs with the same read-only MCP tools and must agree; the desk records the more conservative P(profit) and every veto. UNDERWRITE_SECOND_MODEL, UNDERWRITE_COMMITTEE=review|independent.
• Gate — deterministic Python. Re-quotes every leg through alpaca-py, recomputes max loss itself, enforces 1% of equity per structure, DTE 7–45, bid/ask ≤ 15%, open interest ≥ 100, at most one structure per underlying, four total. Sizes from measured calibration, not from confidence. One structure per underlying, six open at most.
• Executor — places the multi-leg order through the MCP server's place_option_order. What the server returns is recorded verbatim as the executor's claim.
• Auditor — asks the Alpaca CLI, a separate binary with its own HTTP client, what the account actually holds, and reconciles it against the claim. Claimed filled but not filled is a silent failure, and one silent failure freezes sizing at the floor.
• Calibration — Brier score and ECE of stated p_profit against realised outcomes, reported-vs-verified claim accuracy, silent-failure count. Position size unlocks from 0.25% to 1% of equity only when ≥10 structures have resolved with ECE ≤ 0.15, Brier ≤ 0.25 and 99%+ claim accuracy. The metric code is reused from Attest (standard definitions, no LLM anywhere in it).

Two channels, one truth: MCP acts, CLI audits. That is how the hackathon's "MCP or CLI" requirement becomes the point of the product rather than a checkbox.

Run

brew install alpacahq/tap/cli          # audit channel
uv sync                                 # Python 3.12 venv with google-adk, alpaca-py, mcp
cp .env.example .env                    # ALPACA_API_KEY / ALPACA_SECRET_KEY of a FRESH $100k paper account, GOOGLE_API_KEY
uv run underwrite doctor                # CLI, MCP server and alpaca-py all reach the account
uv run underwrite run --dry             # one cycle: audit → strategist → gate, no order
uv run underwrite loop --every 900      # trade until the close; audits every claim; stops when the market closes
uv run underwrite report                # docs/REPORT.md + docs/index.html + docs/ledger.json

The MCP server is launched on demand with uvx alpaca-mcp-server (stdio) — nothing to install.

Strategist model: UNDERWRITE_MODEL is a native Gemini id by default; set it to openrouter/<vendor>/<model> (with OPENROUTER_API_KEY) to route through LiteLLM — the desk did exactly that mid-session on 2026-09-03 when the Gemini key ran out of quota, without stopping.

The ledger

Append-only JSONL under ledger/, one file per record kind: proposals, gate, orders (claims), audits (CLI observations), outcomes. The report is generated from these files and nothing else; every number in it has a line behind it.

Honest limits

• Calibration needs resolved trades. In a 7-day hackathon the desk trades at the floor the whole time — which is the correct behaviour for an agent that has not yet proven it can predict its own outcomes, and the report says so.
• Earnings dates are not checked; the allowlist is liquid index ETFs and mega-caps to keep event risk low.
• Paper trading fills at the NBBO mid-ish; live fills would be worse. The auditor records the actual fill prices so that gap is measurable, not assumed.

Apache-2.0.