Autonomous Solana spot trading agent with hard safety constraints enforced below the LLM.
$30 starting capital. Three positions max. Claude Haiku decides; code enforces the kill switches. Built in public to stress-test the design of Agent Bazaar — running a real autonomous agent with real money is the fastest way to sharpen a protocol meant to coordinate them.
Status: Day 0 — scaffold. Not live. Not funded.
A long-running Node process that:
- Decides every 3 min (± 30s jitter) via Claude Haiku
- Enforces stops, drawdown limits, and daily trade caps every 15s in pure code — no LLM in the path
- Reconciles on-chain wallet state against its own DB every 5 min
- Liquidates everything on
/halt, hard-stop, or unresolvable mismatch - Logs every decision, intent, trade, tick, and error to SQLite
- Reports to Telegram and a Tailscale-only Express dashboard
The whole design philosophy is that execution and risk dominate at $30 — decision quality matters at scale, but at this size you mostly need to not blow up from bugs, slippage, or front-running.
| Rule | Value |
|---|---|
| Starting capital | 30 USDC + ~5 SOL gas |
| Max position size | $5 USDC notional |
| Max concurrent positions | 3 |
| Per-position stop | −40% from cost basis, 2-tick confirm |
| Soft pause | −$8 from peak portfolio value |
| Hard stop | −$13 from peak (full liquidation) |
| Trade cap | 12 / rolling 24h |
| SOL gas floor | 0.015 SOL |
| Slippage budget | ≤ 1.5% one-way at quote time |
| Spot only | No perps, no leverage, no bridging |
Full spec lives in docs/spec.md (added in a later commit).
TypeScript / Node 20+ · Claude Haiku 4.5 (decisions) · Claude Opus 4.7 (daily review) · Jupiter Ultra v3 (execution) · Helius (RPC + tx parsing) · Dexscreener (market data) · SQLite WAL (state) · async-mutex (loop coordination) · Telegram (alerts + control) · Docker on Hetzner CX21 / Frankfurt · Tailscale-only access.
git clone https://github.com/naaku11/coinboi.git
cd coinboi
cp .env.example .env # fill in keys; never commit
npm install
npm run typecheck
npm testDon't run npm start yet — there's nothing to run. Build sequence below.
- Day 0 — scaffold, repo, license
- Day 1 — tx pipeline + Sol CLI wrapper, $1 test swap
- Day 2 — universe filters, signals, portfolio with cost-basis math
- Day 3 — scheduler, decision loop, safety loop, reconciler, halt handler
- Day 4 — observability (SQLite, Telegram, dashboard)
- Day 5 — chaos tests
- Day 6 — paper trade 24h
- Day 7 — fund wallet, go live
Daily updates on Twitter as positions close. Tag: #buildinpublic #solana #aiagents.
MIT — see LICENSE.
If you fork this and run it with real money, you accept that bugs in this code can lose your funds. The author runs it at $30 because at that size the loss is data, not pain. Choose your own size accordingly.