Two strategies — a buy-low-sell-high grid and a breakout trend follower — running as always-on daemons against live market data, with every fill, refit and kill written to a journal you can read. One shared wallet funds them all, capital moves to wherever the opportunity is, and every decision is scored against what actually happened. Paper mode by default: real prices, simulated fills, no exchange keys. Risk caps are enforced on the server, not in the strategy, and an AI agent can tune the grid every hour — but never touch the caps.
Markets range and markets trend; a bot that only knows one of them is wrong half the time. Both engines sit behind the same seam, so a run can change its mind without changing its wallet.
A ladder of buy and sell levels across a price range: geometric or arithmetic spacing, configurable levels and per-level allocation. Buys the dips, sells the bounces, re-anchors itself when the geometry changes and never abandons inventory on a refit. Geometry is set by the agent or runs in mechanical mode: a fixed ±% band, re-centred on a timer.
Long-only entries when price closes above the recent high with the fast average over the slow one; exits on a ratcheting trailing stop sized by volatility with a hard floor. Regime activation is machine-scaled from 4h and 1d ADX and efficiency ratio, with hysteresis so it doesn't flap: the arm deploys its whole slice when the regime turns up and exits cleanly when it fades.
A 1d EMA-cross core position built in staggered tranches, bought on a pullback or after a time limit — the long book beside the trade book, so a grid never sells the last of what you meant to keep.
Flip a live run between grid and trend from the dashboard. The daemon restarts and hands over cleanly: open entries cancelled with partial fills booked, working exits carried as legacy orders, unguarded positions taken over at cost.
A budget pinned to a symbol sits idle while the move happens somewhere else. Here every run draws from one shared wallet, and an allocator moves the slices — inside limits it enforces on itself.
No fixed budget splits. Each symbol declares its grid and trend arms; a trend arm is funded when its regime turns up and released when it fades.
Rebalances the pool toward wherever the opportunity is, and keeps a reserve it refuses to spend — gtbot_pool_reserve_pct is headroom the allocators cannot touch.
The shared-budget invariant is enforced in three layers — when you save, in the allocator, and in the daemon — so a slip in one is caught by the next.
Retire and purge finished runs, seal a wallet, or let the cap follow account value in "use all funds" mode. One wallet, multi-asset seeded, funds everything.
An arm sitting idle names its cause. A silent grid raises an alert. An arm with nothing running has to explain itself to the watchdog, or it pages you.
By default a position is never sold below cost unless the ladder has nothing left to trade with. Liquidating is a decision, not a reflex.
Live market data, fills simulated locally, balances in a simulated wallet. No exchange account, no keys. Every strategy, chart, alert and risk rail behaves exactly as it would with money on the line.
Orders go to the exchange's sandbox with testnet keys. Same code path as live, so what you see here is what you will get there.
Mainnet orders with your own keys. Gated on purpose: the run's own switch and the server's environment flags must all agree; a mismatch is refused at boot, and the daemon fast-fails without credentials. The go-live gate (gtbot-golive) reads the ledgers and passes only when every check does; the account audit (gtbot-audit) checks that the real exchange account backs them, and a finding pages you only if it survives two consecutive audits.
A bot is only as safe as the thing that can say no to it. Every rail below applies identically in paper, testnet and live mode.
Grid geometry is tedious to tune by hand and markets move. So an assistant like Claude does it on a schedule — inside limits it cannot change, with a scoreboard it cannot hide from.
The whole system is drivable by an AI agent over the Model Context Protocol, as a logged-in user: status, the hourly routine brief, grid preview and set, refit proposals, market and backtest, P/L report, decisions and events, budget and allocation mode, create, start, stop, kill, retire and purge a run.
Once an hour an agent takes one compact brief of every run — signal digest, regime and gates, a candidate grid, track record, a "material change" bit — and re-fits the flagged runs' range, levels and deployment, reallocating budget slices where it helps.
Every agent decision is journaled with its reasoning, then scored by the server against what the market actually did: each decision records its clamps and candidate delta and gets a counterfactual verdict computed from intra-bar OHLC. The agent reads its own track record in the next brief and learns from it.
The agent may re-fit geometry and move slices. It cannot raise a risk cap, clear a kill switch or liquidate inventory — those stay human. A dead-man's-switch cron takes over with quantile refits if the routine goes quiet for two hours.
Per-run charts — trades, price history, trend line, grid band, lifecycle markers — holdings, the shared-wallet tile with an inline budget editor, hold / release funds, start / pause / reload per run. A wallet NAV ledger and a P&L report benchmarked against HODL and USDT.
One process per active run, kept alive by a watchdog; market data collected every ten minutes; a nightly backup. No root, no service manager required.
Every fill, refit, switch, halt and kill is an event with an id. Alerts carry the id, so a genuine repeat is distinguishable from a double-send. Regime episodes and engagement tracking show how long each arm was eligible and how much of that time it was actually working. Fills are booked at the price they traded at.
A long-horizon 1d / 1w regime detector with hysteresis and scored 7-day and 30-day outcomes, with an optional daily Telegram digest. A detector, not a predictor: the predictor version was tested and refuted.
Backtest a candidate grid or trend setting from the dashboard or the agent; walk-forward sweep scripts for the parameters that matter.
Runs, orders, events, market data and wallet as documented endpoints with role checks and an audit log — the same surface the dashboard and the MCP server use.
gtbot-golive reads the ledgers and exits 0 only when every check passes; gtbot-audit checks that the real exchange account backs them. The BNB fee float is handled, and exchange switches belong to the host, so a deploy never pushes them.
Going live, incident recovery, alerts, the agent routine's prompt and design, the evidence behind the guardrails — written down, in the repo.
Built on GoatCheese: the admin, API and MCP server are generated from the data model; the strategy engines and daemons are the custom part.
Clone it, run the installer, create a paper run and watch it work. Want a strategy of your own, another exchange, or an agent routine tuned to your rules? Send a brief.
Experimental software for research and paper trading. Nothing here is financial advice; if you connect real funds, you do so at your own risk.
This site uses Google Analytics to count visits. Without your consent it sets no cookies and sends only anonymous signals. Details in the privacy policy.