@hypelens/agent-guard

OpenClaw-native risk sidecar (aka agent-risk) for funded Polymarket operators.

Senpi proved chain-truth exits fix P&L. Polymarket has no native stops. Place skills put risk on; this product keeps exits/risk outside the LLM.

Who it’s for

Funded OpenClaw / agent operators trading Polymarket (optional Hyperliquid watch) who currently lose money because stop-loss, trailing, daily loss, and max exposure live in LLM prompts or cron.

Why place skills aren’t this

| Place skill (polymarket-place, hyperliquid-place, …) | agent-guard | |---|---| | Hero path = enter / bet / order | Hero path = protect / exit on breach | | Positions often from agent memory | Positions only from PM Data API (+ optional HL clearinghouse) | | Risk text in SOUL.md / cron | Hard policy in YAML; agent cannot loosen | | Health = model self-report | Health = CLOB/Data probe + watcher heartbeat | | Alerts via the trading agent | Alerts via file/log + webhook (does not trust the agent) |

This package exposes no place / buy / entry tools.

Install

npm install -g @hypelens/agent-guard
# or
npx @hypelens/agent-guard

Attach a funded wallet (operator)

1. Copy policy:

cp config/policy.example.yaml config/policy.yaml
# edit wallet: "0xYourPolymarketProxyOrFunder"

2. Export env (wallet is enough for discovery + dry-run enforcement):

export AGENT_GUARD_WALLET=0xYourFundedPmWallet   # proxy/funder preferred
export AGENT_GUARD_POLICY=./config/policy.yaml
# optional live protective sells (operator accepts stop sells):
# export AGENT_GUARD_EXIT_PK=0x…                 # NEVER 0x9548…
# then set exits.dryRun: false in YAML

3. Run the watcher beside the trading agent (not inside its prompt loop):

npm run watcher
# → writes state/heartbeat.json + state/guard-alerts.log

4. Wire MCP for OpenClaw / Claude:

{
  "mcpServers": {
    "agent-guard": {
      "command": "npx",
      "args": ["-y", "@hypelens/agent-guard"],
      "env": {
        "AGENT_GUARD_WALLET": "0xYourFundedPmWallet",
        "AGENT_GUARD_POLICY": "/absolute/path/to/policy.yaml"
      }
    }
  }
}

5. Arm: call guard_arm. Disarm requires confirm:true and confirmDisarm:"DISARM".

MCP tools

| Tool | Purpose | |---|---| | guard_status | Floors + health (API probe + heartbeat) — not SOUL | | guard_positions | Data API / clearinghouse discovery | | guard_set_policy | Tighten overlay only (cannot loosen YAML floors) | | guard_arm | Enable enforcement | | guard_disarm | Disable — explicit confirm required | | guard_tick | One-shot discover→evaluate→alert→(dry-run)exit |

Policy (YAML)

See config/policy.example.yaml:

• polymarket.hardStopPct / trailingStopPct / dailyLossLimitUsd / maxExposureUsd
• Agent overlays may only tighten these
• exits.dryRun: true by default (safe adoption)
• Fee stub default free — optional adapter later, only while a position is actively guarded (no SaaS invented here)

HARD rules

1. Never trade/farm 0x9548B8E9554a1968843B3C380431b10996247c88
2. No place-order hero path
3. PM-first; HL clearinghouse is optional discovery/alerts
4. Positions never from agent memory

ClawHub skill

Draft skill at skill/SKILL.md — use-when: funded agent / protect positions / stop loss / trailing. Not place/trade entry.

cd skill/scripts && npm install && node start-mcp.mjs
# separate terminal:
node start-watcher.mjs

Tests

npm test

Policy enforcement unit tests cover hard stop, trailing, daily loss, max exposure, tighten-only overlay, and disarm confirm.

Fee

Default free for adoption. fee.enabled only activates a stub that records “would charge while guarded” — wire a real adapter later. Do not invent SaaS in this package.

License

MIT