Harnesses

Agent-native apps run inside a harness — a host environment that provides the AI agent and displays the app UI side by side.

Claude Code Harness (Local)

  • Open source, ships with @agent-native/harness-claude-code
  • Runs locally — xterm.js terminal on the left, your app iframe on the right
  • Powered by Claude Code CLI via a real PTY (node-pty)
  • Settings panel for launch flags (--dangerously-skip-permissions, --resume, --verbose, custom flags)
  • Restart button to relaunch with new settings
  • Auto-detects when Claude finishes generating and notifies the app
  • Best for: solo development, local testing, open-source projects

Quick start:

# In your agent-native monorepo
pnpm dev:harness

Builder Harness (Cloud)

  • Provided by Builder.io — available at builder.io
  • Runs locally or in the cloud
  • Real-time collaboration — multiple users can watch/interact simultaneously
  • Visual editing capabilities alongside the AI agent
  • Parallel agent execution for faster iteration
  • Best for: teams, production deployments, visual editing, real-time collaboration

Feature Comparison

FeatureClaude Code HarnessBuilder Harness
Local developmentYesYes
Cloud/remoteNoYes
Real-time collaborationNoYes
Visual editingNoYes
Parallel agentsNoYes
Agent chat bridgeYesYes
File watcher (SSE)YesYes
Script systemYesYes
Open sourceYesNo
CostFree (BYOK)Builder.io plan

How It Works

Both harnesses support the same core agent-native protocol:

  1. postMessage bridge — app sends builder.submitChat messages up to the harness
  2. Chat running events — harness sends builder.fusion.chatRunning events down to the app
  3. File watching — SSE endpoint keeps UI in sync when the agent modifies files
  4. Script system — pnpm script <name> dispatches to callable scripts

Your app code is identical regardless of which harness you use.