Docs

Troubleshooting

Problems that are documented, with fixes that are documented. Nothing on this page is a guess.

Symptom

The sidebar shows a red "Backend offline" row after CLAIR was force-closed.

Cause. CLAIR runs its own local backend as a separate process alongside the app window. A crash, a forced end from Task Manager, or a Windows sleep/resume can leave that process stopped while the app shell is still open, or still starting back up.

Fix. If a "Restart backend" button is showing, click it; CLAIR starts a fresh backend process and the row clears on its own once it answers. If the row doesn't clear within about 30 to 45 seconds, or no button appears at all, close CLAIR completely and reopen it.

If you need to force-close CLAIR yourself (Task Manager, a frozen window), end the main CLAIR process rather than just its backend process. Ending only the backend while the app shell keeps running can leave the shell confused about backend state, so it keeps reporting offline even after you reopen the app; a full quit and relaunch clears it.

Symptom

CLAIR shows a message that another Ollama is already running, or the AI engine won't start.

Cause. CLAIR's local AI engine (Ollama) needs a specific local network port. If a separate Ollama installation, most often a standalone Ollama tray app, is already running on this machine, it can be holding that port first.

Fix. Quit Ollama from the system tray (or end it in Task Manager if it doesn't have a tray icon), then restart CLAIR. CLAIR checks for this condition and tells you plainly rather than repeatedly retrying against a port it can't use.

Symptom

Chat feels slow on a machine with an Intel Arc graphics card, and it looks like CLAIR is running on the CPU.

Cause. Intel Arc graphics, including the integrated GPU built into many recent Intel laptop chips, are used through Ollama's own Vulkan graphics backend. By default, Vulkan drops integrated-class graphics devices unless it's explicitly told to use them; CLAIR sets that automatically, but a copy that started once before those settings applied could keep running on CPU silently.

Fix. CLAIR checks, after every launch, whether the loaded AI model is actually resident on the GPU. If it finds the engine stuck on CPU, it restarts the AI engine once on its own and tells you either way. If your specific machine genuinely can't use its GPU for this, CLAIR says so plainly instead of restarting repeatedly. No manual configuration is needed or available for this.

Symptom

Windows shows a blue "Windows protected your PC" screen when you run the installer.

Cause. The installer is not yet digitally signed. Code signing is a paid certificate CLAIR adds after the alpha; during the alpha, every user is personally known and walked through this screen.

Fix. Click More info, then click Run anyway. The installer proceeds normally from there. This does not mean anything is wrong with the file. Full steps: Install CLAIR.

Symptom

First launch appears stuck while downloading the AI model.

Cause. The one-time setup downloads (the AI model, 2 to 17 GB depending on tier, and a library pack of about 1.5 GB) need a working internet connection. A restrictive firewall, VPN, or antivirus product on a managed laptop can block or slow them. The sign-in code and these two downloads are the moments the first launch touches the network.

Fix. Confirm you have an internet connection, temporarily disable a restrictive firewall or VPN if your IT policy allows it, and relaunch CLAIR. The model is fetched once and reused on every later launch; a partial download does not damage anything you've already imported.