Chuchube reads your real repository, plans a change as a graph of tracks, runs each track with its own coding agent in an isolated git worktree, and merges the passing ones back to your branch. You end up with real commits on a branch. No transcript to paste.
Working with a single coding agent turns into a queue. You ask for the frontend, wait. Ask for the backend, wait. Then you ask it to sort out the merge conflict and paste the result into your editor yourself. Context leaks out between every turn, and you're the one holding the thread together.
Chuchube runs the work more like a small team. You write the brief once. It works out which parts can happen at the same time and which depend on each other, then puts a real coding agent on each part.
Three moves: plan, run, merge.
Describe what you want. The architect reads your repository and lays out a graph of tracks, some parallel and some dependent. You review and adjust the whole graph before a single agent starts.
Each track gets its own coding agent in its own git worktree, so they don't step on each other's files. They run at the same time, for real, and you can watch every terminal as it goes.
When a track passes, it merges back into your base branch in dependency order. Real commits land on the branch you picked, ready to review.
A finished run. Five tracks built in parallel, then merged back in order.
- Work is a graph you can see and rearrange, not a linear conversation you scroll through.
- It indexes and reads your actual code before it proposes or writes anything.
- Every track runs in its own worktree, then merges back in dependency order.
- You approve the plan first. Nothing runs until the graph looks right to you.
- Bring the coding CLI you already use. Claude, Codex, Gemini, and Kiro all plug in.
Before the first agent starts, Chuchube parses your repository with tree-sitter, embeds the symbols, and builds a vector index it can query while it plans. Ask for "OAuth login" and it already knows where your routes and auth code live, so the plan points at real files instead of guessing.
Note
Local-first by design. Your source never leaves your machine, and you bring your own provider credentials, so no shared keys ship inside the app.
Each running track is a live terminal. Watch an agent work, type to nudge it mid-run, then let it finish and hand off to the next node. Launch codex, kiro, gemini, or claude per track, whichever you trust for that job.
Live agent terminals. Type to steer a track without stopping the run.
Desktop app for macOS and Windows. It updates itself, so you stay current after the first install. The full download page is at workflow.chuchube.co/download.
curl -fsSL https://workflow.chuchube.co/install.sh | bashTip
The one-line installer is the smoothest path on macOS. It installs the app with no security prompt to click through.
Note
Prefer a file? Download the .dmg from the latest release. The app isn't Apple-notarized yet, so a browser-downloaded .dmg shows a "could not verify" prompt the first time. Clear the quarantine flag once, then open it:
xattr -cr "/Applications/AI Workflow Orchestrator.app"Download the installer from the latest release and run it. Windows may flag an unknown publisher; click More info, then Run anyway.
- macOS on Apple Silicon, or Windows.
- An account or API access for at least one coding agent: Claude, Codex, Gemini, or Kiro.
Chuchube plans with Google Gemini on Vertex AI and pulls code context through MongoDB Atlas Vector Search. The building runs on whichever agent you pick: Claude, OpenAI Codex, Gemini, or Kiro. The desktop app is built with Electron and talks to the cloud planner through a secure proxy, so no API keys are bundled in the download.
Release notes for every version live on the changelog, and a copy is kept in patch-notes/. The current build is v0.1.4, which organizes your work by project.
Grab it from the download page and describe your first change. Not ready to install yet? Join the waitlist and we'll keep you posted.
