Troubleshooting

Download the right build for your machine in one step:

It picks Metal on macOS, Vulkan (or CPU) on Linux, and CUDA for NVIDIA or Vulkan on Windows; force one with --vulkan, --cuda, --cpu or --rocm. If you already have llama.cpp, ArcFlare looks on PATH, in ~/llamacpp/{vulkan,cuda,rocm} and the usual places, or point it at yours:

The current installer doesn't: it installs a private Node LTS into ~/.arcflare/node when yours is missing or older than 18. Run it again, then open a new terminal so the PATH change takes effect.

Run arcflare backend. A build marked --loads but can't see a device. If its kernels list your GPU, the build is fine and the driver isn't exposing the GPU, so update the driver or switch to another backend:

ArcFlare already backs off by itself: the batch first, then the context. To get more room, close other GPU work, use arcflare memory lean, or pick a smaller quant. arcflare fit shows what fits before you load anything.

Pull one with arcflare pull user/repo:QUANT, or set ARCFLARE_MODELS to the folder where your GGUFs already are. Where models are found.

Fixed in current versions: the agent now opens a fresh connection for every request, because llama.cpp's router closes idle ones. Run arcflare update.

A .mcp.json in the working directory starts nothing until you trust it. The agent prints a red ! line saying so. Read the file, then:

If the file changes later, it needs trusting again.

Run arcflare use <harness> --no-launch and read the notes it prints. Your previous config is in <file>.arcflare-bak. ArcFlare won't overwrite a config it can't parse, so fix any syntax error first.

Your torch is a CPU build. NVIDIA: arcflare gen setup <model> --torch cuda. AMD on Windows: borrow a working torch with --torch-from. Setup details.

torchvision was built for a different torch. With --torch-from, install the torchvision that matches the donor's torch into the donor environment, and remove any PyPI torchvision from ArcFlare's venv:

• Check the picture: one whole object, centred, plain background, no hard shadows.
• Raise --octree to 380 and use full Hunyuan3D rather than turbo or TripoSR.
• Try a different --seed.

The session needs to reach the website that relays it. Check which one with arcflare rc relay. The default is http://localhost:3000, which only works on the machine running the site. Point it at your deployment:

The key is valid, but no session is connected with it. Type /rcin a running session. If you rotated the key, paste the new one. If the site is self-hosted on serverless functions, the two halves can't meet, so run it with next start.

It's waiting for approval. A turn sent from the browser asks in the browser, and a turn typed at the terminal asks at the terminal. Unanswered browser prompts are denied after five minutes.

Either you're offline, or your clone has local commits that diverge from main. Offline: use --from. Diverged: git statusin the install folder shows what's in the way.

Checks are off if arcflare update --off was run, if ARCFLARE_NO_UPDATE_CHECK=1 or CI=trueis set, or while you're offline. arcflare update --check checks immediately.