AI Code Generation

The code generator is not a single model — it is a pipeline that picks the right provider for the job, runs it with project-aware tooling, and stitches the output back into your repo. Knowing how it decides helps you steer it.

How the provider is selected

• The model your account selected (Dashboard → model picker).
• Your plan's allowed models.
• The classification of the request — a tiny task may use a faster, cheaper model than a full build.

Project-aware tooling

Most providers run as CLIs that understand a project on disk — they can open files, search the tree, and apply multi-file edits, rather than emitting one blob of text:

• Claude Code CLI — the default for full generations and complex updates.
• Grok CLI — fast, focused tiny tasks.
• Gemini CLI — strong on long, structured outputs.
• Groq — API mode (no CLI required), very low latency.
• Ollama — runs entirely on your machine for privacy.

What one generation does

1. Reads your prompt and the existing project (for updates).
2. Plans the files to create or change.
3. Writes them to disk under your project folder.
4. Installs any new dependencies and saves a version.

What you control

• The model — pick it on the Dashboard, switch it any time.
• The scope — one clear instruction per message keeps edits tight.
• The target — name a file, or @-mention it in the desktop editor.
• The rollback — every generation is versioned, so you can undo.

Common mistakes to avoid

• Ignoring "CLI not found". That provider isn't installed here — switch to Gemini, Groq, or Ollama.
• Huge mixed prompts. The pipeline does best with one goal at a time.
• Not reading the streamed plan. It tells you what's about to change — catch a wrong turn early.

FAQ