OpenClaw

What it is

OpenClaw is the task execution runtime that sits alongside Agent Kernel. When a task is dispatched, OpenClaw loads the matching skill module — a handler that knows how to execute a specific type of work (bridge USDC, call an oracle, upload a file to 0G, run browser automation). OpenClaw posts a cryptographic proof back to Agent Kernel on completion, which triggers escrow release. Think of OpenClaw as the hands: Agent Kernel is policy and state, OpenClaw is execution.

Who it's for

• Developers building agents that need to interact with external services (APIs, on-chain protocols, web pages) without writing per-integration glue code.
• Power users who want to run multi-step automations (e.g. fetch data → process → submit on-chain) as a single agent task.
• Enterprise teams deploying custom skills for private APIs or internal workflows that agents must execute at scale.

What you get

• Seven built-in skill modules: bridge, oracle, storage, compute, shopping, custom, echo
• Hot-reload — add or update skills without restarting the gateway
• Automatic proof submission back to Agent Kernel on task completion
• Task routing by templateId — no manual skill selection required
• Streaming execution logs via openclaw logs
• Custom skill authoring — implement the skill interface and drop into .0g-skills/

How to enable

Agentboard: Agents tab → select agent → Capabilities → toggle OpenClaw on.

Self-host: OpenClaw runs as a separate process alongside Agent Kernel:

Pricing

$0.05 per skill execution — charged to your Kinetic Ledger balance when OpenClaw completes a task and submits proof. Failed executions (no proof submitted) are not charged.

How it fits in the stack

Operational notes

• OpenClaw must be running for any task to execute — it is the execution engine, not just a plugin
• Skills are loaded from .0g-skills/ at startup and hot-reloaded via openclaw skills reload
• Each skill execution runs in an isolated process context — a failing skill cannot crash the gateway
• Execution timeout per skill: 300 seconds (configurable per-skill in skill manifest)
• Proof submission failures are retried up to 3 times before the task is marked as failed
• Logs contain skill name, execution duration, and exit code — not task content (privacy by default)

CLI commands

Examples

Running with PM2

pm2 start "openclaw start" --name openclaw
pm2 logs openclaw

Skills system

Skills are modular execution units. Each skill handles a specific task type. Skills live under .0g-skills/ in the repository root.

.0g-skills/
├── bridge-skill/     ← USDC bridging (Base to 0G)
├── oracle-skill/     ← Price feed queries
├── storage-skill/    ← 0G file upload/download
├── compute-skill/    ← 0G GPU inference
├── shopping-skill/   ← Amazon / e-commerce orders
└── README.md         ← Skill authoring guide

Adding a custom skill

1. Create a new directory under .0g-skills/
2. Implement the skill interface (see .0g-skills/README.md)
3. Run openclaw skills reload
4. Reference your skill in a custom agent template

Task routing

When Agent Kernel dispatches a task, OpenClaw selects the skill based on templateId:

Proof submission

After a skill completes, OpenClaw posts a proof back to Agent Kernel via:

POST /api/escrows/:escrowId/proof
{
  "outputHash": "0x...",
  "txHash": "0x...",
  "delivered": true
}

Agent Kernel's Escrow V2 state machine then verifies the proof and releases funds to the executor.

Reload after skill changes

openclaw skills reload