Official Flame demo
Background AI Coding with Isolated Git Worktrees
See Flame create an isolated task environment, let the agent work without touching the main checkout, and return the result for review.
Parallel AI coding agents
Run AI coding agents in parallel: Claude, Codex & Gemini.
Give each agent an isolated branch and worktree, follow every task from one manager, and review the result before anything lands.
One queue across projects and worktreesSee running, waiting, approval-blocked, failed, and merge-ready work.
The problem
Parallel agents create competing repository state unless each run has a boundary.
Two agents editing one checkout can replace each other's changes, invalidate the index, and run tests against a mixed working tree. Even when the files do not overlap, shared ports, generated output, local databases, and environment state can make one agent's result depend on another run.
The coordination problem continues after generation. Developers need to know which task is running, blocked on approval, failed, ready for review, or unable to merge, then inspect the exact diff before deciding what lands.
How Flame solves it
Give every agent an isolated run and an explicit landing decision.
The task stays connected to its project, branch, worktree, transcript, approvals, verification, and final diff.
- 01
Split work by ownership
Choose tasks that can be reviewed independently and state which files, service, or outcome each agent owns.
- 02
Create an isolated worktree
Give every implementation task its own branch and Git worktree instead of sharing the main checkout.
- 03
Start the right coding agent
Choose Claude Code, OpenAI Codex, or a Gemini-powered agent for each task.
- 04
Watch the queue
Track running, waiting, approval-blocked, failed, scheduled, and merge-ready tasks in Agents Manager.
- 05
Review before landing
Inspect the transcript and diff, run verification, then merge, push, continue, or discard the worktree.
Review queue
Agent completion is not merge approval.
A successful agent run only means the process stopped without reporting failure. It does not prove that the implementation matches the task, preserves surrounding behavior, or passes the repository's checks.
Open the task transcript and changed-file list, inspect the diff, run the relevant tests, and decide whether to merge, push for review, continue the conversation, or discard the worktree.
Merge-blocked tasks remain visible in Agents Manager instead of disappearing into an old terminal session.
Review before landingRead the actual code change and verify it in repository context.
Clear responsibility
Flame manages the workflow. You make the engineering decisions.
Flame handles
- Agent launches in the correct project or worktree
- Task state, transcripts, approvals, schedules, and filters
- Worktree initialization and selected local-file copying
- Diff review and merge, push, continue, or discard actions
You decide
- Which tasks are independent enough to run together
- Which model and permission level fits each job
- How ports, databases, and external services are isolated
- Whether the tested result is ready to land
Product walkthrough
Watch an agent work in an isolated Git worktree.
FAQ
Questions developers ask before changing their workflow
Run the queue without surrendering the repo.
Download Flame and supervise parallel coding agents from task creation to reviewed merge.
Free to use on macOS, Windows, and Linux.