#ACP x CC CLI

1 messages · Page 1 of 1 (latest)

shrewd ivy
#

@fallen ravine Can claude code (with Max plan) run by openclaw through ACP?

idle sunBOT
coral gyroBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> âś… Mark Solution

idle sunBOT
fallen ravineBOT
#

Yep, that should work.

OpenClaw can run Claude Code through ACP, and Anthropic Claude CLI / subscription auth is supported again, including Claude Code use with Pro/Max plans.

Main gotchas:

  • Claude auth must already exist on the host running OpenClaw
  • first /acp spawn claude may need to fetch the Claude ACP adapter
  • ACP is non-interactive, so Claude permission prompts can fail unless ACP permissions are configured appropriately

Good starting point:

  • /acp spawn claude --bind here
  • or /acp spawn claude --thread auto

Docs:

#

If you want, I can help you check the exact setup. If it’s not working already, paste:

  • openclaw --version
  • /acp doctor output
shrewd ivy
#

If ACP is non-interactive then what is the difference between claude --print and ACP?

fallen ravineBOT
#

Non-interactive here just means no live TTY/UI for permission prompts. It does not mean ACP is just claude --print.

Roughly:

  • claude --print

    • one raw CLI invocation
    • prompt in, text out
    • no OpenClaw /acp controls
    • no thread/conversation binding
    • no ACP session resume / background-task tracking

  • ACP

    • OpenClaw runs Claude Code as a managed harness session
    • you get /acp spawn, /acp status, /acp steer, /acp cancel, /acp close
    • can bind the session to this thread/chat
    • supports resume, persistent session state, background task tracking, cwd, etc.
#

So the better mental model is:

  • claude --print = “run Claude once and read stdout”
  • ACP = “OpenClaw supervises a durable Claude Code session”

The catch is just that ACP sessions are headless, so if Claude Code wants to ask “approve this write/exec?”, there’s no terminal prompt to answer it.

Docs:

If you want, I can also give the practical “when should I use ACP vs claude --print?” version.

shrewd ivy
#

Is there any other better option?

fallen ravineBOT
#

Yeah, but “better” depends on what you want.

My honest ranking:

  1. Best overall OpenClaw experience: native Anthropic provider with API key

    • most predictable
    • cleaner tool/runtime behavior
    • what I’d recommend for regular OpenClaw use

  2. Best if you specifically want Claude Code inside chat/thread workflows: ACP

    • persistent session
    • /acp controls
    • thread binding, resume, background tasks
    • more moving parts than native provider

  3. Best if you specifically want to reuse your Max plan with minimal fuss: Claude CLI backend / claude -p

    • simplest path
#
  • good for straightforward prompt → answer flows
    • less powerful than ACP

So, short version:

  • want reliability → native Anthropic provider
  • want Claude Code-as-a-harness → ACP
  • want use Max plan as simply as possible → claude -p / CLI backend

Docs:

If you want, I can give you the one I’d choose for your exact use case if you tell me whether you want:

  • normal assistant/chat use
  • coding agent in a repo
#
  • cheapest way to use Max through OpenClaw
shrewd ivy
#

Can you confirm if using ACP is 100% legitimate for claude code (with Max plan)?

fallen ravineBOT
#

I can’t honestly say 100% in a legal/contractual sense.

What I can confirm from the OpenClaw docs is:

  • OpenClaw says Anthropic staff told them OpenClaw-style Claude CLI usage is allowed again
  • OpenClaw therefore treats Claude CLI reuse / subscription auth as sanctioned for this integration
  • the docs also link Anthropic’s own article for Claude Code with Pro/Max

So from the project’s guidance, the answer is basically: yes, this is treated as legitimate/supported.

What I cannot confirm is:

  • an explicit Anthropic policy line that says, word for word, “OpenClaw ACP with Claude Code on Max is 100% guaranteed allowed forever”

So my careful answer is:

Practically yes, but I would not call it a 100% legal guarantee.

shrewd ivy
#

Can CC knobs like plan, ultraplan, effort, Adaptive Thinking, etc., can be managed through ACP?

fallen ravineBOT
#

Not as clean first-class controls, no.

What OpenClaw ACP does document/manage directly is mostly:

  • model
  • permissions / approval policy
  • timeout
  • cwd
  • set-mode
  • generic /acp set <key> <value>

So:

  • if by plan you mean an ACP runtime mode, maybe yes via /acp set-mode plan
  • if you mean Claude Code specific knobs like ultraplan, effort, Adaptive Thinking, etc., I do not see those exposed as standard OpenClaw ACP controls right now

The important bit is that ACP supports a generic config setter, but whether a given key actually works depends on the backend + Claude adapter. If the adapter doesn’t support it, OpenClaw should return an unsupported-control style error.