Tara LLM API Key Saver

TaraApiKeySaver is the one-time setup step for the whole Tara pack, and its signature move is that nothing visibly happens when it runs. It's an output node with no outputs. You drop it on an empty canvas, pick a provider, paste a key, hit Queue - and that's the save. The README is upfront that confirmation "will work" someday; as of the alpha it just writes the file silently. It feels broken the first time. It isn't.

What it does

Three inputs:

• provider - dropdown of groq, openai, together. The key lands under a matching field (GROQ_API_KEY, OPENAI_API_KEY, ...) in API_KEYS.json.
• api_key - the key itself, pasted in.
• temporary - default off. Off writes into the pack's own folder (.../ComfyUI-Tara-LLM-Integration/py/API_KEYS.json). On writes to /tmp instead.

The mechanism is boring in a good way: it reads the existing JSON if present, merges in just your provider's key, and writes it back with orjson. Save a Groq key today, an OpenAI key next week - both coexist in the same file. Then TaraApiKeyLoader (or TaraPresetLLMConfig with use_loader on) reads them back later, and your actual generation workflows never contain a key.

The catch: it's plaintext

Let's be clear about what "securely saves" in the README means. It means conveniently, not encrypted. This is a JSON file with your credentials in plain text, in a custom node folder - and ComfyUI custom nodes are unsandboxed code running as you, the same structural fact behind the LLMVISION malware incident. The README's own caution is the important part: on a hosted ComfyUI (Runpod and friends), temporary writes to /tmp, which Linux clears on reboot but which on a shared box is shared - another user's save can overwrite your key. The author's fallback, and the one I'd actually use on any cloud box, is to skip this node entirely and pass the key through a text/primitive node that never touches disk.

The workflow

1. Add TaraApiKeySaver to an empty graph
2. provider = groq (or openai), paste your key
3. Queue → the write happens, no visible output
4. Clear the canvas
5. Build your real workflow; TaraApiKeyLoader fetches the key

Troubleshooting

• Loader says "API key not found" - you queued the Saver, right? Queueing is the action; editing the node without running does nothing. Also check that the temporary flags match between Saver and Loader.
• Someone else's key appears - you're on a shared host with temporary on. That's the README's warning made real; use a primitive node instead.
• No error, but no file - check the pack's py/ folder for API_KEYS.json and confirm it's not being written to a virtualized/overlay path you can't see.

Install is the pack-standard two-liner (git clone https://github.com/ronniebasak/ComfyUI-Tara-LLM-Integration into custom_nodes, restart), or Manager → Install via Git URL. Deps are just orjson and openai - no models, no heavy downloads.