cyboflow Home Features Docs Blog Changelog Download

The manual.

How to set up Cyboflow, and what each part of the app does: the flows, the board, the review queue, sessions, design mode, visual verification, the assistant, and models and permissions. And how findings from one run change the next.

01What Cyboflow is

Cyboflow is a macOS and Windows desktop app for running AI coding agents in parallel against one project. By default each session works in its own isolated git worktree, on its own branch, so several agents move at once without touching your working copy.

A quick session can opt out and work in place in your project checkout instead. Flow runs are always worktree-backed.

What a flow run needs from a human is routed into one place: Human review, which is also the page the app opens on. Two things are approved elsewhere: the assistant, in its rail on the right of the app, confirms its own proposals in place, and a design is approved on the design canvas.

Nothing a flow produces reaches your default branch until you merge it.

02The mental model

The pieces, and one loop.

  • Projects are local git repos you add to the app. Point it at a folder that is not a repo yet and it runs git init rather than refusing the folder.
  • Sessions hold the work. By default a session owns an isolated git worktree. A quick session can opt into working in place in the project checkout. Runs nest inside a session, flow runs and chat turns alike, and you can keep several sessions open at once.
  • Flows are multi-step agent workflows. Six ship in the box: Launch, Planner, Sprint, Ship, Compound, and Verify Setup. Five appear in the launch wizard, and Verify Setup starts from the Verify Queue or a project’s overview. You can edit any of them, or build your own.
  • The backlog is a shared board of ideas, epics, and tasks. Launch and Planner fill it, Sprint drains it, Ship does both inside one run, and Compound can add follow-up tasks. Sprint runs each task in its own lane, and a task marked as human work never gets one.
  • The review queue collects what needs a human, in five kinds: permission requests, decisions, human tasks, notifications, and findings. Most of it comes from flow runs, though a quick session can put items there too. Findings are the exception to the list: they are triaged in Insights, and one reaches the queue only when the agent filed it as blocking.
  • Design mode is a fullscreen session mode. It binds to one idea, and you refine a prototype there and approve it onto that idea.

The left rail carries five destinations, each with its own subtitle in the app.

Human reviewPending approvals. The landing page, headed Human review queue.
Task backlogPlanning pipeline. Ideas, epics, and tasks on one board.
InsightsFindings · stats · quality.
WorkflowsFlows & agents. The gallery, the step editor, and the agents a step binds to.
Verify QueueVisual verification. Run-book health per project, and where a Verify Setup run starts.

The assistant rail sits on the right of those surfaces.

The loop starts with an idea. Capture one yourself, or interview a brand-new project with Launch, which produces a project brief and turns it into an ordered set of ideas. Planner takes an idea and returns approved, ready tasks, with an optional design pass that produces a prototype and an architecture design first.

Sprint executes ready tasks in the session’s worktree, at most five task lanes at once. Ship plans and executes the whole arc in one continuous run, from an idea stub to integrated tasks, so Sprint is the flow you reach for when the tasks already exist.

Then you read the diff and merge it. Compound mines merged work for durable lessons and applies the ones you approve, which is how findings raised during one run change the next one.

03The guides

12 pages, in the order the rail lists them.