Claw Harness

Testing framework for OpenClaw bots. Spin up real agent instances, load skills and personas, drive multi-turn prompts, and capture results.

Can a real AI agent, given only your skill.md, figure out how to use your site?

Unlike API-level test harnesses, Claw Harness tests the full agent experience end-to-end — skill comprehension, API discovery, tool usage, and multi-agent interaction.

Quick Start

npm install claw-harness

Prerequisites:

• Node.js >= 22
• OpenClaw installed (npm install -g openclaw@latest)
• ANTHROPIC_API_KEY set in environment

Run an example

# Try the hello world scenario (no target app needed)
claw-harness run examples/hello-world.yaml

# Or scaffold your own
claw-harness init my-test
claw-harness run my-test.yaml

See examples/ for more scenarios you can run immediately.

Programmatic API

import { ClawHarness } from 'claw-harness'

const bench = new ClawHarness({ mode: 'local' })

const bot = bench.bot('alpha', {
  preset: 'default',
  skills: [{ url: 'http://localhost:3000/skill.md', name: 'my-app' }],
  soulMd: 'You are a friendly bot.',
})

await bench.start()
const response = await bot.send('Register yourself on the platform')
console.log(response.text)
await bench.stop()

Scenario Format

Scenarios are YAML files that define bots and a sequence of steps:

name: "Chat Test"

bots:
  alpha:
    preset: default
    model: anthropic/claude-haiku-4-5-20251001
    soul_md: presets/personas/friendly.md
    skills:
      - url: "http://localhost:3000/skill.md"
        name: target-app

  beta:
    preset: default
    soul_md: presets/personas/curious.md
    skills:
      - url: "http://localhost:3000/skill.md"
        name: target-app

steps:
  # Serial steps
  - bot: alpha
    prompt: "Read the skill docs and register yourself."
    timeout: 60s

  - bot: beta
    prompt: "Register yourself on the platform."
    timeout: 60s

  # Parallel steps
  - parallel:
      - bot: alpha
        prompt: "Join a lounge and introduce yourself."
      - bot: beta
        prompt: "Find a lounge with another bot and join."

  # Repeat block
  - repeat: 3
    interval: 15s
    steps:
      - bot: alpha
        prompt: "Check for new messages and respond."
        timeout: 30s
      - bot: beta
        prompt: "Check for new messages and respond."
        timeout: 30s

CLI

claw-harness run <scenario.yaml> [options]   Run a test scenario
claw-harness init [name]                     Scaffold a new scenario
claw-harness presets                         List available presets

Options:
  --model <model>       Override model for all bots
  --reporter <format>   Output format: console (default) | json

Presets

Claw Harness ships with presets for common configurations:

Configs — merged into each bot's openclaw.json:

• default — Full tools, Haiku model
• minimal — Restricted tools, lower cost

Personas — soul.md templates that shape bot personality:

• friendly — Outgoing, asks follow-up questions
• curious — Thoughtful, explores ideas deeply
• terse — Brief, technical, to the point

Examples

The examples/ directory has self-contained scenarios you can run right away — no target app required:

Scenario	What it shows
hello-world.yaml	Simplest possible scenario — one bot, one prompt
assertions.yaml	contains, not_contains, and matches assertions
two-bots.yaml	Two bots with different personas, parallel execution
with-cleanup.yaml	after: cleanup steps that run even on failure
secret-menu.yaml	Load a custom SKILL.md and verify the bot follows it

How It Works

Each bot gets full isolation:

1. Workspace — A dedicated profile directory (~/.openclaw-claw-harness-<id>/) with its own openclaw.json, USER.md, and skills
2. Gateway — Its own OpenClaw gateway process on a dedicated port
3. Communication — Prompts sent via the OpenAI-compatible HTTP API (/v1/chat/completions)

Bots have no shared state. They interact only through the target application, just like real users would.

Development

git clone https://github.com/dasconnor/claw-harness.git
cd claw-harness
npm install
npm test          # Run tests (vitest)
npm run build     # Build to dist/
npm run lint      # Type check

License

MIT