Vito-bc/crypto-orchestra: Multi-agent crypto research and paper-trading system with deterministic risk gates and byte-verified backtests. Python · Claude API · Coinbase Advanced. · GitHub

This is a completed negative-result research program with a reproducible evidence harness, not a live trading bot. It asked one question — does any directional crypto mechanism this account can afford to run have a demonstrable edge on Coinbase — and it answers no, on four separate and differently-grounded lines of evidence, plus a fifth line (funding carry) that is priced but not yet worth starting. Every headline number below is a byte-identical regeneration from a pinned environment, including the numbers that say the strategy does not work.

Seven Claude sub-agents and a deterministic risk engine exist to run a directional momentum mechanism (ZEC-USD, paper/shadow) if one were ever authorized. Over 2026, four separate lines of evidence were opened against that possibility, each closed on its own grounds, plus a fifth priced and left open. All four closures and the carry pricing are recorded in docs/trial_registry.md, which is the authoritative source for every number below.

The reusable result. The perpetuals closure derives a floor identity that applies to any directional rule on a given asset, not just the two it tested:

floor_annual = 1.645 · SD_bar · sqrt(bars in position per year / Y)

Trades-per-year times mean-hold-in-bars is bars-in-position-per-year, so the trade count cancels out of the annual floor exactly. Trading a directional rule more often cannot lower the smallest edge that history could tell apart from zero — it only multiplies the fee bill. This is scale-free: it depends on volatility and time in market, not on account size, so a larger balance does not reopen any of the lines above.

What would reopen a line, and what would not. Two recorded conditions: Coinbase Financial Markets' own funding, now recorded forward because the exchange has no funding-history endpoint, sustaining above the carry break-even for longer than a cycle; or a venue with materially lower fees becoming reachable by a New York resident. A bigger account balance is explicitly not one of them.

What runs today: a daily execution-cost probe (optionally credential-scoped view-only for its fee-tier reading) and the credential-free hourly CFM funding monitor described in docs/operations/cfm_funding_monitor.md. What does not run: the seven-agent pipeline, any LLM call on a schedule, and any order path. The mechanism below is what would run if a line were ever reopened and authorized — it is documented, tested, and currently idle.

The claim this project actually stands on: the committed research artifacts regenerate byte-for-byte, on a pinned interpreter and pinned result-determining libraries, from a local candle cache.

Both this and the cheaper --verify-code run in CI, fed by credential-free public hydration. Exact source-file hashes, the declared pins for the whole numerical dependency closure, and the installed environment form one content-addressed provenance identity; input identity is a window-scoped logical OHLCV hash. The interpreter is pinned to 3.13.5 exactly — regenerating under anything else is refused rather than silently producing different numbers. See docs/research/.

flowchart LR
  subgraph A["7 sub-agents · claude-haiku-4-5 · 5-worker pool"]
    direction TB
    A1["technical — RSI, MACD, BB, EMA"]
    A2["macro — 4h regime, veto power"]
    A3["sentiment — Fear and Greed"]
    A4["whale — funding, BTC dominance"]
    A5["risk — ATR stop and target"]
    A6["news — asset headlines"]
    A7["breakout — price structure"]
  end

  S["signal scanner — momentum breakout"] --> A
  A --> O{{"orchestrator — claude-sonnet-4-6"}}
  O --> G["deterministic risk gate — 6 entry filters, circuit breakers"]
  G -->|allowed| L["limit order — simulated, DRY_RUN"]
  G -->|blocked| H["HOLD"]

If ever run continuously: every 60 minutes the pipeline closes positions that hit stop / target / max-hold, reconciles pending limit orders, runs the scanner, and — only if the scanner produces a signal — spends tokens on the agents. The orchestrator's BUY still has to clear the risk gate before an order is written. None of this is scheduled today; see "Research program" above for what actually runs.

The agents propose; deterministic code disposes. Every gate below is plain Python, testable without an LLM, and fails closed — an unreadable input blocks entry rather than waving it through.

Sizing is LIVE_BALANCE_USD x position_size_pct, defined once in pipeline/sizing.py and defaulting to $100. The circuit breakers measure against the same baseline. It is not an aggregate spending cap: concurrent positions can add up to more than that baseline. TRADE_SIZE_PCT defaults to 2% and cannot exceed the deterministic 12% per-trade ceiling.

The code contains a live path for auditability, but this project does not authorize its use and the Quick Start does not enable it. Activation would require, in order: a repaired evaluation harness, a pre-registered strategy with an acceptance rule fixed in advance, a forward shadow trial that passes it, and an explicit human decision recorded in docs/trial_registry.md. None of those conditions is met, and the research program above found no candidate that would meet them. LIVE_BALANCE_USD=100 would remain only the per-trade sizing baseline — it is not a portfolio cap and not evidence that money is at risk today.

Requires Python 3.13.5 (exact — the research tooling refuses other versions). No exchange credentials are needed for anything below.

python -m venv venv
venv\Scripts\activate                    # Windows
pip install -r requirements.txt
copy .env.example .env                   # fill in ANTHROPIC_API_KEY

python pipeline/runner.py ZEC-USD        # one paper run, one asset
python pipeline/scheduler.py             # continuous loop, every 60 min
streamlit run app.py -- --demo           # dashboard on the synthetic UI fixture

Research, offline and credential-free:

python backtesting/hydrate_research_data.py      # public Coinbase candles
python backtesting/research_runner.py --verify   # byte-identical regeneration
python -m pytest -q                              # hermetic: no network, no .env

The --demo dashboard renders demo/ — a hand-authored synthetic fixture for exercising the UI. It is not a backtest and not a track record. Real results are in docs/research/.

agents/          7 specialist sub-agents + orchestrator
pipeline/        runner (risk engine), limit orders, position tracker,
                 scheduler, sizing, dashboards, Telegram summaries
exchange/        all Coinbase calls, isolated for auditing
backtesting/     signal scanner, research runner, walk-forward, Monte Carlo,
                 the perpetuals feasibility gate, the carry scoping tool
tools/           price data, support/resistance levels, market positioning
schemas/         Pydantic contracts shared by every agent
docs/research/   research index, artifacts, manifests, closure documents
docs/            trial registry (the authoritative record), ADRs, task briefs
tests/           1,300+ tests — hermetic, network denied at the socket layer

LIVE_BALANCE_USD is unrelated to START_BALANCE in backtesting/ — that is a simulation convention so historical replays report readable dollars, not a cap.

• .env and cdp_api_key.json (Coinbase ECDSA key) are git-ignored — never commit either.
• DRY_RUN=true is the default; no real order is placed without an explicit opt-in.
• Coinbase API keys must have can_transfer=false; preflight treats withdrawal rights as a blocking error.
• Authenticated order calls are isolated in exchange/coinbase_client.py; public candle hydration lives in exchange/coinbase_candles.py.
• Vulnerability reports: SECURITY.md.

See CONTRIBUTING.md. The short version: research claims need a pre-registered trial ID in docs/trial_registry.md before the scan, not after, and every PR must keep lint, tests and research-verify green.

MIT.