A coding agent that proves its work
Plan, edit, check, test, review. Helix Code works in your real project, streams every step as it happens, and does not call a change done until your project's own checks pass.
Watch it think and type
Replies stream in as they are written; in testing the first words arrive in about a second and a half. Shell commands and test runs stream live in a terminal panel, so a long install never looks stuck.
Keep what's right, undo what isn't
After every turn, open the review: each file's changes hunk by hunk. Put back a single change, a whole file, or the entire turn. If the file changed since you looked, Helix refuses rather than reverting the wrong lines.
Everything a careful engineer does
Checks every edit
Each file is checked the moment it is written: syntax, undefined names, broken JSON. Like red squiggles, for the agent.
Repo map
Ranks the files the rest of the project depends on, weighted to your task, so big projects need no blind reading.
Test-first fixes
For a bug, it writes a test that fails, fixes the code, and shows the test going green. Fixes that are proven.
A done gate
Your typecheck, lint and tests must pass before "done", and a self-review reads the diff for anything that breaks it.
Missions
Hand over a goal; it works in its own git worktree until the checks pass, then hands you a branch to merge.
Your stronger model
Add your own Claude or OpenAI-compatible key and it plans and reviews with it. Edits stay fast on the default.
Give it the change you keep putting off.
Helix Code is built into Helix for Windows.