A diagnostic map
Start with the symptom
Use the message shown by MRCO AI, then test the smallest part of the stack that could explain it.
| Symptom | What to check first | Next step |
|---|---|---|
| App stops on startup | Error message, available disk space and runtime downloads. | Review %LOCALAPPDATA%\AiDesktop\crash.log. |
| Online provider fails | Endpoint path, API key, exact model name and proxy settings. | Use Test and Fetch; check the final chat-completions URL. |
| Model replies but skips tools | Tool Calling support and tool availability for this request. | Run the model's Tool Calling test; try a compatible Instruct model. |
| Local model will not start | GGUF path, llama-server runtime, RAM/VRAM and GPU settings. | Try a smaller model or temporarily disable GPU offload. |
| Vision produces no result | Vision model and projector installed, target window identifiable. | Install/check assets in Features and try a focused screen question. |
| Microphone does not work | Whisper installed, recording device selected and Windows permission. | Test recording access, then retry speech input. |
| No spoken reply | Speak Replies enabled and a compatible SAPI voice installed. | Check whether the Edge TTS fallback can reach the network. |
| Semantic search returns nothing | Embedding model installed and the target folder indexed. | Try a more specific query and verify file formats are supported. |
| File will not enter Knowledge | File exists, read access, size and supported format. | Note that scanned image-only PDFs may not contain extractable text. |
| Browser automation fails | Edge/Chrome available, Playwright setup and target site access. | Run the browser test before retrying the task. |
Debug with discipline
Change one variable at a time.
Repeatedly switching providers, models and modes at once hides the cause. Preserve the exact error and validate each layer separately.
- ✓Record the exact prompt, selected mode and model.
- ✓Check whether basic chat works before testing tools.
- ✓Test Tool Calling and browser/vision integrations separately.
- ✓Keep secrets out of logs, screenshots and public bug reports.
ISOLATION SEQUENCE
01 Does the app start?
02 Does basic chat respond?
03 Is the selected provider healthy?
04 Can the model call the required tool?
05 Does the target integration work alone?
06 Can the result be independently verified?
OUTPUT One reproducible issue, one isolated cause.
01 Does the app start?
02 Does basic chat respond?
03 Is the selected provider healthy?
04 Can the model call the required tool?
05 Does the target integration work alone?
06 Can the result be independently verified?
OUTPUT One reproducible issue, one isolated cause.