Binance Agent OS
A trading co-pilot on Binance Agent OS. It scans the market, scores setups, and enforces a fixed risk budget on every position. Seven risk gates run before any proposal, so a trade that breaks your rules never reaches you. You confirm every real order.
Try it
Click a command to see how Sentinel responds. This is a scripted walkthrough of the real decision flow. Connected to live Binance data through the Agent OS MCP.
The spine
No stage is skippable. The risk gate sits before the proposal, so a trade that breaks your rules is never suggested.
Opportunity score
Every score shows its full breakdown. A high score flags a setup worth considering. Whether it can be traded is a separate decision, made by the risk gates.
The risk engine
Position size drives nothing. Risk drives size, volatility sets the stop, the stop sets the size. Each gate shows its real number, not a verdict you have to trust. This is the SOL proposal clearing every check.
The moment that matters
Ask for something reckless and Sentinel refuses before it proposes, names the gate that failed, and gives you the largest size that would pass every check. A clear no, and a usable yes, in the same answer.
Under the hood
A system-prompt agent, not a script. It binds to whatever tools the live MCP exposes, so it survives API changes. No keys in the repo. Account access is brokered through the authenticated agentic sub-account.
You natural language + commands │ ▼ ┌───────────────────────────────────────────────┐ │ SENTINEL │ │ system prompt = analyst + risk officer │ │ │ │ Scan ▸ Analyze ▸ Score(0-100) ▸ │ │ Risk Gate(7) ▸ Propose ▸ Confirm ▸ │ │ Execute (paper by default) │ └──────┬──────────────────────────────┬─────────┘ │ reads live data │ local, no keys ▼ ▼ ┌─────────────────────────┐ ┌────────────────────────┐ │ Binance Agent OS MCP │ │ memory/ + config/ │ │ ticker · book · klines │ │ risk profile │ │ funding · balances · │ │ watchlist │ │ positions · orders │ │ paper trades + last │ │ (agentic sub-account) │ │ proposal │ └─────────────────────────┘ └────────────────────────┘
Controls
Natural language works too. Sentinel maps intent to the closest command and confirms briefly.
scan marketRank the watchlist by opportunity score.
analyze <SYM>Full multi-timeframe read + score breakdown.
risk dashboardEquity, exposure, daily budget, live gates.
portfolioBalances, open positions, unrealized PnL.
propose <SYM>Run the workflow and emit a sized proposal.
confirm paperLog the last proposal as a paper trade.
disable safety modeHuman-only. Live trades still need confirm.
set risk <profile>Conservative / Balanced / Aggressive.
Get it running
Add the Agent OS server to your agent client and authorize the sub-account.
Load AGENT.md as the system prompt. That single file is the whole agent.
Type scan market. Safety Mode is on by default, so it reads live data and trades only on your confirmation.