Frequently asked questions

A self-hosted agentic AI platform for developers. You describe a goal; it inspects your repository, plans, edits code in an isolated workspace, runs your real tests, and hands you a diff with the evidence.

No, not by design. Inference runs on a model you host (vLLM, SGLang or Ollama), and execution happens in a sandbox on your infrastructure. Nothing is pushed to a remote for you.

Any Qwen coding-capable instruct model served through an OpenAI-compatible API, such as Qwen2.5-Coder-32B-Instruct on vLLM or qwen2.5-coder:7b on Ollama. The model sits behind an adapter, so it can be swapped without touching orchestration.

A 7B coder model runs on a modern laptop with 16 GB of RAM through Ollama. Larger models such as 32B need a GPU server; benchmark your choice for tool-call reliability, speed and memory before committing.

No. git.push is denied in every mode for the MVP. You review the diff and apply it yourself.

Supervised asks you before every file write and protected action. Bounded autonomous runs routine, permitted actions on its own, but network access, dependency installs, destructive operations, pushes and merges still need your approval. Both are capped by budgets.

It can't loop forever: steps, tool calls, wall time and retries are all capped. When a limit is hit the task stops with a partial report, and recoverable tasks can be resumed with an extended budget.

Not yet. The console can drive the control-plane API, which inspects repositories read-only and produces validated plans; executing patches in an isolated sandbox arrives in Sprint 3. See the roadmap for what's next.

A license hasn't been chosen yet. Check the repository before redistributing, and confirm the license of the model you deploy separately.