Questions

The ones people ask before they run the first command.

Questions

Do I need a GPU?

For the scratch lane, no: about a million parameters trains on a CPU in minutes to hours depending on the corpus. A consumer card makes it quick. The fine-tune lane, when it lands, wants eight to twelve gigabytes of VRAM for the default 1.5B base, four for the 0.5B, and more for 3B.

Does anything leave my machine?

No. Indexing, training, asking and tool execution all run locally. The site only ever stores a recipe or a bundle record you explicitly save to your account, and never your files or your weights.

Who decides a tool call?

In the scratch lane, a rule planner reads the question and picks a tool; the model does not emit calls, and the trace labels each step planner routed. In the fine-tune lane the model emits the call itself, in a schema the backend validates before executing. Both lanes fence tools to the roots you allow and gate tiers B and C behind confirmation.

What is the fine-tune lane and is it available?

A LoRA over a small open instruct model, Qwen2.5-1.5B-Instruct by default, trained on your curated dataset plus tool-call transcripts so the model itself emits schema-valid calls. It is the roadmap: finetune.py is not in the ported tree yet, and every page that mentions it says so.

Which base models and licences?

Qwen2.5 0.5B and 1.5B under Apache-2.0, Qwen2.5 3B under the Qwen research licence, SmolLM2 1.7B under Apache-2.0, and Llama 3.2 1B under the Llama community licence, which you accept on Hugging Face before download. The bundle page shows the licence of the base it was built on.

Who owns the bundle?

You do, in every lane. A bundle you train locally never touches our servers. A bundle trained on the hosted lane, when it exists, is yours to download at any time and is deleted when you say so.

Why is the demo sometimes a recording?

The demo bundle runs on a local sidecar. When none is up, the page shows a transcript recorded from the repo's own eval suite and marks it recorded so it is never mistaken for a live answer.