Agents can act.
Authority stays bounded.
Binance AgentGuard is the accountable action layer for Binance AI Agents: it turns a natural-language goal into a bounded Spot intent, grants authority only inside policy, independently verifies reality, and seals every outcome in a tamper-evident receipt.
Judge in 60 seconds
Run a goal
Click a preset below to watch natural language become a structured intent and complete Agent trace.
Inspect the boundary
Policy authorization and post-trade verification are deterministic—not delegated to the model.
Verify real evidence
Open the Binance order bundle or run one command to recompute its evidence hash.
One Agent loop, three accountable outcomes
Authority exercised safely
Order filled, identity bound, notional inside budget, and slippage inside policy.
Authority not granted
The request violates policy or lacks confirmation. No Binance order is submitted.
Authority revoked after mismatch
Execution returned an unsafe result. The task is isolated and cannot continue.
Executable policy
The AI may propose an action, but it cannot rewrite these boundaries or approve itself. Judges can change the sandbox policy in the console and immediately inspect the resulting decision.
Three layers of evidence
Each layer answers a different judge question, and every layer is labelled so simulation is never presented as a live trade.
Binance Spot order
Previously authorized BNBUSDT market BUY. Order 12512896470, FILLED for 0.007 BNB / 4.81047 USDT, independently queried and hash-bound.
AI planner + Binance MCP
An actual model generated the intent and observed a live Binance quote. The guard returned BLOCKED because transaction confirmation was absent. No order was submitted.
Every failure branch
The public console deterministically demonstrates VERIFIED, BLOCKED, and FROZEN without an account, wallet, API key, or OAuth token.
Real Binance Spot evidence
12512896470Verify evidence in this browser
Load the public Binance evidence bundle and independently recompute its SHA-256 with Web Crypto.
Recorded: 039b9c5c006b66ecd2c243990cb8b453c2d52c6e3b9c01b199fb890ee03b07f3This is a Binance Spot order record—not a public blockchain transaction hash.
Let the Agent act inside boundaries
Give the Agent a natural-language goal, adjust its sandbox authority, and inspect every deterministic decision. The simulator never touches an account.
1 · User goal
2 · Adjustable policy
All controls are local sandbox inputs. No wallet, API key, OAuth token, or real order is used.
3 · Agent trace
Waiting for a goal.
4 · Independent verification matrix
| Check | Expected | Observed | Result |
|---|---|---|---|
| Run a goal to inspect each control. | |||
5 · Receipt
{}Why Binance
This is not a generic chatbot with an exchange logo. Binance supplies the execution environment and authenticated source of truth that the guard is designed around.
Agentic account boundary
The Agent runs against a dedicated Binance Agentic account context instead of embedding exchange credentials in application code.
Official MCP tool surface
Market reads and Spot execution are mediated through Binance Agentic MCP, giving the policy layer a narrow, auditable gateway.
Exchange-native verification
AgentGuard independently re-queries Binance order and trade records rather than trusting the Agent’s claim that execution succeeded.
Run in under five minutes
npm install
npm run demo:agent
npm run judge:checkThe judge command validates code, nine policy/verification tests, both public live evidence bundles, all three Agent states, and required public-site signals.
Live MCP read
Deterministic planner
npm run live:run -- --read-onlyReads the Agentic Spot account and a fresh BNBUSDT quote, then fails closed without confirmation.
Actual AI planner
npm run agent:liveUses a local Codex model to construct the plan. The model still cannot authorize or submit an order.