Good catch tracking it down to AVX. The honest answer is that a non-AVX build isn't something this fork can ship. The add-on installs the official @anthropic-ai/claude-code package and runs it on the Node.js runtime from the base image, and the AVX requirement comes from those upstream pieces, not from anything in the add-on. I don't compile either of them, so there's no switch here to turn AVX off.
That said, it depends on where you're running it. If this is a Proxmox (or other) VM, the usual cause is the VM's CPU type being left at the default (kvm64/qemu64), which hides AVX from the guest even when the physical host CPU has it. In the VM's hardware settings, set Processor > Type to host, reboot the VM, and AVX gets passed through. That clears up the large majority of "no AVX" reports.
If the physical CPU genuinely has no AVX (older hardware, some low-power chips), there's no software workaround, and the realistic option is to run the add-on on a machine that has AVX, or run Claude Code on a small separate box and point at it.
The most common cause: the terminal wraps that very long URL across two lines, and when you select across the line break your browser quietly slips in a space or newline, or grabs only part of it. That breaks the permission list partway through.
Try this, in order:
Before touching the link, zoom your browser OUT with Ctrl+- (Cmd+- on Mac) until the whole URL sits on ONE unbroken line in the terminal. This is the single most reliable fix. A one-line URL cannot pick up a hidden break.
Only once it is on one line, click the link (or triple-click that single line to select it, copy, and paste into the address bar). Do not click while it is still wrapped, because the terminal treats each wrapped row as its own link and the top one opens a cut-off URL, which is probably what bit you.
If it still wraps, lower terminal_font_size in the app config (smaller font means more characters per line, so the URL is far likelier to fit) and/or maximize the browser window. Then retry step 2.
Leave session_persistence on its default (false) for first-time login. That keeps tmux out of the way so normal copy/paste just works.
That clears it up for most people. If a clean one-line URL still gives you "Unknown scope," that points away from copying and toward the Claude Code version that got bundled at install time. In the terminal run claude --version, then npm install -g @anthropic-ai/claude-code@latest, restart, and try again.
Before diving in, would someone mind giving an overview of exactly what access this gives to Claude Code? E.g. full API, everything in /config, something less or more?
I'm figuring git out and have been using Claude code on some other projects.
I'm thinking about how the .storage folder has all of the API keys for any integration that's been configured.
I'm less concerned about it botching an automation than exposing something that shouldn't. Not that I think Anthropic has any use for my keys, but it just doesn't seem like a best practice.
I guess I don't fully understand how effective any boundaries you can set with Claude Code or similar are. E.g. .claudeignore
I was thinking about it... Partially, I just want to understand these better, and make a conscious decision about exposure vs accidentally seeing something get accessed that I didn't expect.
I’m actually wishing that home assistant was structured a little differently. I want to go the other direction than discussed in that issue. I’m fine with giving a LLM access to most of my yaml and configuration files, but .storage and some other files/folder have secrets that I wouldn’t want to give. Because .storage is in /config, it seems I can’t give fs access without risking it get into that folder.