Analytics

A unified view of every Sentinel service — receipts written, policy compiled, bench coverage, and the agent surface guarded by the gate.

Merkle computing…

3-minute judge path

Show that Sentinel stops unsafe autonomous execution before it reaches Bitget Agent Hub.

Start with the adversarial call, show the sealed receipt, then back it with the 20-case benchmark and MCP integration surface.

  1. 1Run prompt injectionCompromised agent asks for DOGE and 50x leverage.
  2. 2Run transfer attemptUnsafe fund movement is blocked and sealed.
  3. 3Open Bench + MCPShow 20/20 adversarial coverage and the proxy surface.
Bench pass rate +100%
— 20 adversarial scenarios sealed
Sealed receipts Merkle
— root computing…
Violations caught harmful intents
— block + repair verdicts

Verdict composition by category

attack surface · sorted by severity
Block Modify Allow

Intercept mix

verdict ratio
— Intercepted
  • Block——
  • Modify——
  • Allow——

Service health

drill in

Recent receipts

latest bench run
ScenarioToolSymbolVerdictReceipt
loading…
—Bench pass
—Blocked
—Repaired
—Allowed
DefaultActive policy
PaperExecution mode
Send a Bitget Agent Hub call through Sentinel:

Receipt stream

live
TimeToolSymbolVerdictReasonReceipt
Click a trigger above to send an audited tool call through the Sentinel gate.

Active mandate

Describe the agent's mandate in natural language. Sentinel compiles it via Qwen (with a deterministic regex fallback) into a frozen policy object.

Compile the mandate to see the deterministic policy that the gate will enforce.

Violation catalog

Every code below is a structured reason the gate can emit. Receipts carry the codes verbatim so reviewers can replay decisions.

    —Total pass
    —Blocked
    —Repaired
    —Allowed
    —Categories
    —Pass rate

    Adversarial scenarios

    20 cases · deterministic Merkle root

    Stdio JSON-RPC proxy

    The Sentinel MCP proxy wraps Bitget Agent Hub over stdio. Every tools/call for a write-class tool is intercepted, evaluated by the gate, and either passed through to Agent Hub, repaired before forwarding, or blocked with a receipt.

    Tools that mutate state (orders, leverage, transfers, withdrawals) are routed through the gate. Read-only perception skills pass through transparently.

    $ npm run mcp:proxy [Sentinel] MCP proxy listening on stdio [Sentinel] Bitget Agent Hub: upstream connected [Sentinel] Guarding 7 write tools

    Surface

    Tools currently routed through the Sentinel gate:

    • futures_place_orderguarded
    • futures_modify_orderguarded
    • futures_cancel_orderguarded
    • futures_set_leverageguarded
    • spot_place_orderguarded
    • transferguarded
    • withdrawguarded
    • market_data_*passthrough
    • account_balancepassthrough

    Bitget Agent Hub

    Sentinel integrates with Bitget Agent Hub through a thin adapter. Tool calls arrive as MCP requests; the adapter normalizes them into a Sentinel intent, runs the gate, and either executes via the live Agent Hub or replies with a sealed block receipt.

    Five perception skills (macro analyst, market intel, news briefing, sentiment analyst, technical analysis) read through the proxy without being gated. Execution-class tools are always gated.

    // src/adapters/agent-hub.js const WRITE_TOOLS = new Set([ "futures_place_order", "futures_modify_order", "futures_cancel_order", "futures_set_leverage", "spot_place_order", "transfer", "withdraw" ]); export async function guardedToolCall(toolCall, ctx) { const intent = toolCallToIntent(toolCall); const decision = await ctx.gate.handle(intent); if (decision.verdict === "block") return blockResponse(decision); return ctx.hub.invoke(decision.intent); }

    Perception skills

    Read-only context layered into the agent prompt — never gated, never sealed.

    • macro-analystread-through
    • market-intelread-through
    • news-briefingread-through
    • sentiment-analystread-through
    • technical-analysisread-through

    Repository

    github.com/dmetagame/sentinel-flight-recorder — the deterministic gate, MCP proxy, Agent Hub adapter, Qwen integration, benchmark suite, and this dashboard.

    Continuous integration runs the unit suite plus the 20-scenario bench on every push.

    CI status Vercel

    Layout

    • src/core/risk.jsgate
    • src/core/policy.jspolicy
    • src/core/merkle.jssealing
    • src/adapters/agent-hub.jsadapter
    • src/adapters/qwen.jsllm
    • src/mcp-proxy.jsstdio
    • src/bench/scenarios.jsbench
    • public/index.htmldashboard