#Hello, I’m completely new to Openclaw. I don’t understand how to make my agent run independently.

1 messages · Page 1 of 1 (latest)

next heron
#

It keeps asking me for confirmation; it says it’s doing something but doesn’t actually do it. Here’s an excerpt from my openclaw.json

  "agents": {
    "defaults": {
      "workspace": "/home/mycky/.openclaw/workspace",
      "heartbeat": {
        "every": "15m",
        "prompt": "Vérifie tes tâches en cours et continue l'exécution si une étape est restée en suspens."
      },
      "model": {
        "primary": "lmstudio/google/gemma-4-e4b",
        "fallbacks": [
          "google/gemini-3-flash"
        ]
      },
      "thinkingDefault": "adaptive"
    },
    "list": [
      {
        "id": "perso",
        "workspace": "~/.openclaw/agents/perso/workspace",
        "name": "Ellie",
        "identity": {
          "name": "Ellie",
          "emoji": "👩‍💻"
        },
        "model": {
          "primary": "lmstudio/google/gemma-4-e4b",
          "fallbacks": [
            "google/gemini-3-flash"
          ]
        },
        "tools": {
          "profile": "full",
          "alsoAllow": [
            "browser",
            "tavily_search",
            "tavily_extract",
            "memory",
            "cron",
            "files"
          ]
        }
      },
      {
        "id": "main",
        "name": "Main",
        "model": {
          "primary": "lmstudio/google/gemma-4-e4b",
          "fallbacks": [
            "google/gemini-3-flash"
          ]
        },
        "tools": {
          "profile": "full",
          "alsoAllow": [
            "browser",
            "message",
            "nodes",
            "tavily_search",
            "tavily_extract"
          ]
        }
      }
    ]
  },

  "models": {
    "mode": "merge",
    "providers": {
      "lmstudio": {
        "baseUrl": "http://localhost:1234/v1",
        "apiKey": "lmstudio",
        "api": "openai-responses",
        "models": [
          {
            "id": "google/gemma-4-e4b",
            "name": "google/gemma-4-e4b",
            "reasoning": false,
            "input": [
              "text"
            ],
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            },
            "contextWindow": 131072,
            "maxTokens": 4096
          }
        ]
      },
      "google": {
        "baseUrl": "https://generativelanguage.googleapis.com/v1beta",
        "apiKey": "",
        "api": "google-generative-ai",
        "models": [
          {
            "id": "gemini-3-flash",
            "name": "gemini-3-flash",
            "input": [
              "text"
            ],
            "contextWindow": 1048576
          }
        ]
      }
    }
  },

I'm using WSL, and LM Studio Gemma 4

tranquil wigeonBOT
lapis otterBOT
#

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

tranquil wigeonBOT
ancient galleonBOT
#

This looks like two separate things mixed together:

  1. Heartbeat is not “autonomy mode”
    Heartbeat is just a periodic nudge. For real autonomous work, use standing orders + cron.
    Docs: https://docs.openclaw.ai/automation

  2. Your local model is probably the real blocker
    LM Studio + local Gemma often does this bad pattern: it narrates actions, asks for confirmation, or pretends to use tools instead of actually doing them. The docs specifically call out Gemma/local backends as failing on normal agent turns.
    Docs: https://docs.openclaw.ai/gateway/local-models
    Troubleshooting: https://docs.openclaw.ai/gateway/troubleshooting

Also, one important detail:

  • your fallback is effectively not configured, because google.apiKey is empty.

First things to paste

# 
openclaw --version
openclaw status
openclaw models status
curl http://127.0.0.1:1234/v1/models

That last one matters extra on WSL. If OpenClaw runs in WSL and LM Studio runs on Windows, localhost may not be the right address from the WSL side.

Fast diagnostic test

Temporarily add this to your LM Studio model entry:

"compat": {
  "supportsTools": false
# 
}

If the weird behavior changes immediately, that confirms the model/backend is choking on tool schemas.

My recommendation

For reliable autonomous behavior:

  • use a hosted primary model
  • keep LM Studio as a fallback
  • put the rules in AGENTS.md
  • trigger them with cron, not only heartbeat

If you paste:

  • openclaw --version
  • one openclaw logs --follow excerpt after reproducing
#
  • and your tools.exec section if the “confirmation” is an actual approval prompt

I can tell you exactly which part is broken.

rugged ginkgo
#

Well man, my advice is go reinstall again. Delete everything, the bootstrap is the key to get you started and clean. Go brainstorm yourself and write a long prompt, what you want to achieve and what you want your new buddy to do, and how you want it to handle any "decisions". Basically it will not do things you didnt permit it in the start. So the first "prompt" to it will build its soul. You can always use a CHATGPT or CLAUDE to help you brainstorm on what to prompt to OpenClaw.

2nd advice, based on what you want your new buddy do. Go think what are the things a usual human would need. If you want to build a house, you will need engineers, architects, and etc. Same thing go to CHATGPT or CLAUDE ask them to help you brainstorm. Yes building house is just random example I pull up. THis part will help you build up your sub-agents, and next would be your workflow. Build the workflow of the task you want him do. This will determine and you need put in every details on each steps so that it will continue the task till the next. If you are using CLAUDE most probably a lot of the steps it will provide you the prompt for human intervention. So you may want to edit that part.

And most probably in your "soul.md" there is some restriction you didnt intentionally set. So it will always ask for confirmation.

ancient galleonBOT
#

Small correction, I would not start with a reinstall here.

A few things from the docs:

Also, in this config:

  • primary = lmstudio/google/gemma-4-e4b
#
  • fallback = google/gemini-3-flash
  • but google.apiKey is empty

So the fallback is basically not ready.

And local Gemma/LM Studio can absolutely do the “I’m doing it...” / “please confirm...” / “pretend tool use” thing when it is not handling the full agent/tool surface well. The local-model troubleshooting docs call out this class of problem directly:
https://docs.openclaw.ai/gateway/local-models
https://docs.openclaw.ai/gateway/troubleshooting

Best next checks

openclaw --version
openclaw status
openclaw models status
curl http://127.0.0.1:1234/v1/models
#

That last one should be run from inside WSL. If OpenClaw is in WSL and LM Studio is on Windows, localhost might be wrong from the WSL side.

If the “confirmation” is an actual OpenClaw approval prompt, then that is a different path, and the next thing to inspect is tools.exec. But if it is the model itself asking in chat, I’d bet on the model/backend first.

next heron
ancient galleonBOT
#

Not as the first thing, no.

What I’d do after install is:

  1. Finish onboarding / first run
  2. Check basic health first
    openclaw --version
openclaw status
openclaw models status
  3. From WSL, verify LM Studio is actually reachable
    curl http://127.0.0.1:1234/v1/models

Then, yes, it’s worth a quick look at: