29 replies

ACP agent configuration

✅Solved

So i told my openclaw assitant to create agents for me in my discord channel using ACP i wanted claude, gemini and codex it did the follow config below. is it correct and are these the best models to use for the channel or is there a bestter way to do this, take note that i have the various clis installed claude code cli, gemini cli, codex cli. i assumed it will connect directly to the CLIs.

{
"id": "claude",
"name": "Claude Coding Agent",
"workspace": "/Users/.openclaw/workspace-coding",
"model": {
"primary": "anthropic/claude-sonnet-4-6"
}
},
{
"id": "codex",
"name": "Codex Coding Agent",
"workspace": "/Users/.openclaw/workspace-coding",
"model": {
"primary": "openai/gpt-5.4"
}
},
{
"id": "gemini",
"name": "Gemini Coding Agent",
"workspace": "/Users/.openclaw/workspace-coding",
"model": {
"primary": "google/gemini-3.1-flash-lite-preview"
}
}
]

Solution

That config is “normal OpenClaw agents that call model APIs”, not ACP/CLI-backed agents.

1) Will it connect to your installed CLIs?

No — having Claude Code / Gemini CLI / Codex CLI installed does not auto-wire them.
To use those CLIs, you need ACP runtime (via the ACP backend plugin, typically

acpx

acpx

) and an agent (or binding) whose

runtime.type

runtime.type

"acp"

"acp"

. See: https://docs.openclaw.ai/tools/acp-agents

Your snippet only sets:

model: { primary: "provider/model" }

model: { primary: "provider/model" }

…which means “use this provider over HTTP”, not “run the CLI”.

2) Is the config “correct” as-is?

It’s syntactically plausible for API-based agents, but a few red flags: