Skip to content

Instantly share code, notes, and snippets.

@goldmar
Last active March 12, 2026 02:31
Show Gist options
  • Select an option

  • Save goldmar/fb9c24330177c2c0822a2b7185957fd3 to your computer and use it in GitHub Desktop.

Select an option

Save goldmar/fb9c24330177c2c0822a2b7185957fd3 to your computer and use it in GitHub Desktop.
openclaw-code-agent vs built-in ACP: full coding agent orchestration for OpenClaw

openclaw-code-agent: Claude Code orchestration for OpenClaw

The dedicated OpenClaw plugin for Claude Code — more powerful than the built-in ACP integration. https://github.com/goldmar/openclaw-code-agent (dropping open source soon)


Built-in ACP vs. openclaw-code-agent

Feature Built-in ACP openclaw-code-agent
Launch Claude Code sessions
Multi-turn sessions ✅ Fully resumable
Plan mode (propose before coding) ✅ Native support
Approve / revise plans inline approve=true or send feedback
Session persistence across restarts ✅ Serialized to disk
Parallel sessions ✅ Unlimited concurrent sessions
Stream live output mid-session agent_output anytime
Auto-respond rules ✅ Define autonomy level per session
Cost tracking per session ✅ USD, per-session
Session history & stats agent_sessions, agent_stats
Resume a previous session by ID ✅ Full history preserved
Fork a session to explore alternatives

The difference in practice

The built-in ACP is a relay bridge — it connects your OpenClaw agent to Claude Code CLI. Great for simple one-shot tasks.

openclaw-code-agent is a full orchestration layer:

You: Build a REST API for todos

Alice → agent_launch(prompt="...", permission_mode="plan")

[Claude Code proposes plan: 5 files, REST endpoints, PostgreSQL]

Alice: Here's the plan — want any changes?
You: Add rate limiting

Alice → agent_respond(session, "Add rate limiting to all endpoints")
Alice → agent_respond(session, approve=true)  // once revised

[Claude Code implements — silently, in the background]
[You get a notification when it's done]

No babysitting. No copy-pasting. Full async, full control.


Tool surface

agent_launch     — start a session (background)
agent_respond    — reply mid-session or approve a plan
agent_output     — stream live output
agent_sessions   — list all active/recent sessions
agent_kill       — terminate a session
agent_stats      — usage metrics and costs

Going open source shortly → https://github.com/goldmar/openclaw-code-agent

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment