Portrait of Michael Limberger

Michael Limberger

Need me? Email mike@limberger.ca

AI

Troubleshooting

When it breaks

Everything that can go wrong, why it goes wrong, and how to fix it. This section is organized by symptom, because when something breaks you usually see the symptom first, not the root cause.

"0/200k tokens" or TUI shows no response

This is the most reported issue. You type a message, hit enter, and nothing happens. The token counter stays at zero. The cursor blinks. Silence.

Root cause 1: Missing or wrong api field in the config. 

The "api" field must be "ollama" (for native mode) or
        "openai-responses" (for compatibility mode). If it's missing or
        set to something else, OpenClaw sends requests in a format that
        Ollama doesn't understand. Ollama ignores the malformed request
        and sends nothing back. No error on either side.

        Fix: Add "api": "ollama" to your provider config.

Root cause 2: Mismatched baseUrl and api mode. 

With api: "ollama", the baseUrl must be http://127.0.0.1:11434
        (no /v1). With api: "openai-responses", it must be
        http://127.0.0.1:11434/v1. Cross them and you get silence.

        Fix: Match them correctly. See the configuration section.

Root cause 3: Ollama is not running.

Show me

curl http://localhost:11434/

        If that doesn't return "Ollama is running", start Ollama:

        ollama serve

        Or launch the Ollama app from Applications.

Root cause 4: The model is not loaded.

Show me

ollama list | grep qwen2.5-coder

        If the model doesn't appear, it wasn't pulled successfully.
        Re-pull it:

        ollama pull qwen2.5-coder:14b

"No API key found for provider ollama"

OpenClaw's provider initialization needs a non-empty API key string even though Ollama ignores it.

Fix: set OLLAMA_API_KEY=ollama-local in your environment, or add "apiKey": "ollama-local" to the Ollama provider block in openclaw.json. The value is ignored by Ollama but required by OpenClaw.

Gateway port conflict (EADDRINUSE)

Error: listen EADDRINUSE: address already in use :::18789

Something else is using port 18789. Find out what: 

lsof -i :18789

Common culprits:

A leftover clawdbot-gateway or moltbot-gateway service from before
        the project was renamed. Check:

        launchctl list | grep claw

        Another OpenClaw instance. Kill it:

        pkill -f openclaw

        Or just use a different port:

        OPENCLAW_GATEWAY_PORT=19001 openclaw gateway run --bind loopback

"unauthorized: gateway token missing" (WebSocket 1008)

The WebSocket connection to the gateway was rejected because no auth token was provided, or the token was wrong.

Fix:

Show me

openclaw doctor --generate-gateway-token
        openclaw dashboard --no-open

The second command prints the authenticated URL with the token embedded. Copy that URL into your browser. Since version 2026.2.19+, OpenClaw auto-generates and persists a gateway token at startup, so this mostly hits older versions or corrupted configs.

Model too large for RAM

When the model exceeds available memory, macOS starts swapping to disk. Symptoms: responses take minutes instead of seconds, fans spin up, the whole system feels stuck. Ollama does not crash. It just gets painfully slow.

How to diagnose: open Activity Monitor, click the Memory tab. If the Memory Pressure graph is yellow or red, you are swapping.

Fixes, in order of preference:

Show me

1. Close memory-hungry apps. Each Chrome tab uses 100-300MB. Close
           everything you're not actively using.

        2. Reduce context window. Edit your Modelfile or config to set
           num_ctx to 4096 or 8192 instead of the default.

        3. Use a smaller quantization. Q3_K_M is about 7.3GB for 14B,
           compared to 9GB for Q4_K_M:

           ollama pull qwen2.5-coder:14b-q3_K_M

        4. Use a smaller model entirely. Qwen3 8B at roughly 5GB fits
           easily on 16GB machines:

           ollama pull qwen3:8b

Agent describes actions instead of doing them

You ask the agent to list files and it says "I would run the ls command..." instead of actually running ls and showing results.

That means tool calling is not working. The model is generating conversational text instead of structured tool-call responses.

Root causes:

Wrong model. DeepSeek-R1 and some older models don't support tool
        calling. Stick with qwen2.5-coder:14b.

        Wrong api mode. Make sure "api" is set to "ollama" (not missing,
        not something else).

        reasoning: true in model config. Set it to false for Ollama models.
        When true, OpenClaw sends a "developer" role message that confuses
        the model.

macOS permissions aren't taking effect

You toggled Full Disk Access or Accessibility in System Settings, but OpenClaw still cannot access files or use certain tools.

The fix is embarrassingly simple: quit Terminal completely and reopen it. macOS does not apply permission changes to already-running terminal sessions.

Closing a tab is not enough. Fully quit Terminal (Cmd+Q) and relaunch it.

This catches almost everyone at least once. It is a long debugging trap because nothing in the error messages suggests "restart Terminal."

Ollama crashes mid-conversation

Ollama can crash from memory pressure, macOS putting the process to sleep, or bugs in specific model inference paths.

What to do:

1. OpenClaw's gateway keeps session state independently. Your
           conversation history is preserved in
           ~/.openclaw/agents/<agentId>/sessions/

        2. Restart Ollama:

           pkill ollama
           ollama serve &

        3. Re-warm the model:

           ollama run qwen2.5-coder:14b "ping" --verbose

        4. Send your next message in OpenClaw. It reconnects automatically.

Cloud fallback is happening

You configured Ollama, but the logs show requests going to Anthropic or OpenAI.

Diagnosis:

Show me

openclaw logs --follow

Send a test message, then check logs for "anthropic" or "openai". 

env | grep -i "anthropic\|openai\|openrouter\|gemini"

If any cloud API keys are set in your environment, OpenClaw may use them instead of Ollama regardless of your JSON config. The priority system puts cloud providers first.

Fix: unset the environment variables:

unset ANTHROPIC_API_KEY
        unset OPENAI_API_KEY

And remove them from your shell profile (~/.zshrc, ~/.bashrc) if they are set there permanently. 

+----------------------------------------------------------+
|  THE NUCLEAR OPTION                                      |
|                                                          |
|  If everything is broken beyond diagnosis:               |
|                                                          |
|    rm -rf ~/.openclaw                                    |
|    openclaw onboard --install-daemon                     |
|    # Re-enter Ollama config manually                     |
|                                                          |
|  For Ollama specifically:                                |
|                                                          |
|    pkill ollama                                          |
|    ollama serve &                                        |
|    ollama run qwen2.5-coder:14b "test"                   |
+----------------------------------------------------------+

The full diagnostic command

When in doubt, run the doctor:

Show me

openclaw doctor --fix

That checks config validity, provider connectivity, gateway status, dependency versions, file permissions, and known issues. The --fix flag auto-remediates what it can. Start here when something breaks and you are not sure what.