NickVanzo/run-arena-agent

Run Arena Agent — Autonomous Trading Loop

Start the autonomous trading loop that queries pool state, asks the LLM for a trading decision, and submits intents on-chain every N seconds.

Prerequisites

• AGENT_PRIVATE_KEY set in .env (the wallet must be registered as an agent)
• Agent registered — run register-agent.mjs first if needed
• AGENTS.md present in the current working directory (ai-plugin/)
• LLM access — one of the following:
  • OpenClaw gateway running locally (OPENCLAW_GATEWAY_TOKEN set), or
  • OG_COMPUTE_API_KEY set for 0G Compute direct access

Parsing Interval from User Messages

| User says | intervalMs | |-----------|-----------| | "every 2 minutes" / default | 120000 | | "every 30 seconds" | 30000 | | "every 5 minutes" | 300000 | | "every 10 minutes" | 600000 |

Convert to milliseconds and pass as the first argument.

Command

# Default: 2-minute interval, OpenClaw gateway at localhost:3000
node skills/run-arena-agent/arena-loop.mjs

# Custom interval (e.g. 30 seconds)
node skills/run-arena-agent/arena-loop.mjs 30000

# Custom interval + custom gateway URL + token
node skills/run-arena-agent/arena-loop.mjs 60000 http://127.0.0.1:3000 <token>

With explicit env vars

AGENT_PRIVATE_KEY=0x... \
OPENCLAW_GATEWAY_TOKEN=<token> \
node skills/run-arena-agent/arena-loop.mjs 120000

Using 0G Compute directly (no gateway)

AGENT_PRIVATE_KEY=0x... \
OG_COMPUTE_API_KEY=<key> \
node skills/run-arena-agent/arena-loop.mjs 120000

What It Does Each Epoch

1. Query pool state — calls the Subgraph MCP server to get current price, current tick, nearby tick depth, and the agent's open position.
2. Call LLM — sends pool state to the LLM with the agent's strategy as context (via OpenClaw or 0G Compute).
3. Parse decision — extracts a JSON object with action (open/close/rebalance/hold) and optional amountUSDC, tickLower, tickUpper, reason fields.
4. Submit intent on-chain — if action is not hold, ABI-encodes the intent params and calls AgentManager.submitIntent(agentId, actionType, params).
5. Sleep — waits intervalMs before the next epoch.

How to Stop

Send SIGINT (Ctrl+C) or SIGTERM. The loop exits cleanly after the current epoch finishes.

LLM Routing

OPENCLAW_GATEWAY_TOKEN set?
  YES → POST /v1/chat/completions to OpenClaw gateway
        (gateway injects AGENTS.md as system context automatically)
  NO  → OG_COMPUTE_API_KEY set?
          YES → POST to 0G Compute directly
                (reads AGENTS.md from cwd as system prompt)
          NO  → Error — set at least one of the above

The OpenClaw gateway is preferred because it handles model selection and automatically injects the agent's strategy document.