Skip to content
Sirius IDE Docs

AI Agent overview

The Sirius agent is a tool-using LLM client that lives inside the IDE. It is the closest equivalent to Cursor / Windsurf / Zed’s ACP — but the API key belongs to us, you buy a token quota.

What it can do

  • Read any file inside the open folder.
  • Edit files by proposing diffs that you approve.
  • Run terminal commands (build, test, git) with explicit confirmation for anything destructive.
  • Search project-wide using indexed symbols.
  • Reason over up to ~100k tokens of context, routing between cheap models for trivial queries and stronger models (o3, Opus 4.7) for hard reasoning.

What it can’t do

  • Access the internet on its own (no curl, no pip install without your y/N).
  • Read or write outside the project folder.
  • See your other tabs, browser history, or other applications.

The tool surface is intentionally narrow. You can audit every call in the agent panel’s Actions tab.

How models are picked

There’s a small router inside the backend. By default it sends:

  • Trivial completions and summaries → cheap model (e.g. gpt-4.1-mini).
  • Refactoring and reasoning → mid-tier (gpt-5.2-codex).
  • Hard bugs, planning → o3 or claude-opus-4.7.
  • The “Max quality” button (Plus & Pro) → always the top model, ~2–3× the token cost on your quota.

We continuously tweak the router based on COGS / quality logs. You never see the model name unless you ask — the agent panel has a Show model toggle for diagnostics.