Draw your trade on the chart, get a validated bot back

What it does

You sketch your market read directly onto the chart — trendlines, levels, fibs — and an agent reads your drawings, the visible range, your indicators, and a screenshot to propose setups with entry, stop, target, and invalidation. Pick one and it generates a Freqtrade strategy, validates it, backtests it, and deploys it live on Hyperliquid with bracket orders and a time-boxed auto-stop. It’s a Next.js app with an embedded Postgres, so a self-hosted copy runs with no database provisioning and no login.

The interesting bit

The real product isn’t the code generation — it’s the validator between “the model wrote it” and “it’s trading your funds.” Rejected strategies get their complaints fed back to the model, which rewrites until it passes. Every rule is a scar: no shift(-N) (future-peeking backtests), leverage-scaled stops, startup_candle_count sized against lookbacks so indicators aren’t NaN live. One rule exists because a strategy raised on every candle for ten hours while its entry condition hit seven times.

Key highlights

• Agent reads your actual chart drawings and screenshots, and draws back — levels, zones, channels, fibs
• Deterministic safety validation with an automatic repair loop before any strategy runs
• Runs fully self-hosted: keys stay local, no telemetry by default, every prompt editable
• Fresh clone works on the bundled Lightweight Charts; TradingView Advanced Charts is free but not redistributable, so it’s opt-in
• E2E tests drive the whole app with real model calls but stop short of the two irreversible actions: deploying and withdrawing
• Also agent-friendly: hand a coding agent SKILL.md and it can trade without the UI

Caveats

• It places real orders with real funds by design; the README itself says start on testnet or with money you’d shrug at
• The validator catches known mistakes, not unknown ones — you’re still expected to read what the agent writes
• Self-hosted withdrawals stop one hop short of fully automated (documented in docs/withdrawals.md)

Verdict

Worth a look for anyone building or running LLM-generated trading strategies — the validation pipeline alone is a useful reference for that whole genre. If you want a passive bot with no judgment in the loop, this isn’t it: you’re the analyst, the agent is the (heavily checked) implementer.