Questions,answered plainly.
What Mirven is, how it runs many coding agents on one Unity project, and what it will not do.
What is Mirven?
Mirven is an AI game studio for Unity. It is a desktop app for Windows and macOS that plans the work, runs Claude Code and Codex agents on your Unity project side by side, has every job checked by a model that did not write it, and merges code only when you approve. It runs on your own PC or server with your own Claude Code or Codex account.
How do I run several Claude Code agents on one Unity project at the same time?
Give each script job its own Git worktree, keep scene and asset work in one lane at a time, and write down which folders each job will touch before it starts. Mirven does all three for you: Astra, its planner, splits your request into jobs with named folders and a finish line, script jobs build in separate worktrees, scene and asset jobs take turns, and a job that names a folder a running job already named waits its turn.
Can Claude Code and Codex work on the same project at once?
Yes. In Mirven, Claude Code and Codex agents build in parallel lanes on the same Unity project, each on its own job. The review of each job comes from a different model than the one that wrote it.
Why does my multi-agent setup stall?
Most multi-agent setups route every report through one manager model. Its context fills with everyone else's news until it forgets, repeats itself and stalls. Mirven has no manager model: Astra writes the plan once, and each agent carries only its own job and the notes it needs.
Why do agents break Unity scenes?
Unity scenes, prefabs and many assets are large serialized files that do not merge cleanly, so two agents editing the same scene in parallel produce conflicts. Mirven runs script jobs in parallel worktrees and makes scene and asset jobs take turns in the project itself.
How is Mirven different from running Claude Code subagents or worktrees myself?
Claude Code subagents and worktrees give you the parts. Mirven adds the studio around them: a written plan with a finish line for every job, folder checks before a job starts, Claude Code and Codex in one queue, review by a different model, real Unity builds as checks, project memory, and a merge only when you say yes.
Does Mirven send my code to its servers?
No. Mirven runs on hardware you control. Your agents send what they work on to the model provider you chose, Anthropic for Claude Code or OpenAI for Codex. Your code never passes through Mirven.
Which AI models does Mirven use? Do I need an API key?
Mirven uses your own Claude Code or Codex sign-in. It does not resell model access, and you pay your provider directly.
How much does Mirven cost?
The private beta is free. You pay only your own model usage. Mirven shows the cost of every job when the provider reports it, and says it is unknown instead of showing $0 when the provider does not.
Does Mirven merge code automatically?
No. Automatic merging is off, and stays off unless you turn it on. You review each result and its run log, then decide.
How does Mirven check the agents' work?
Every job goes through the checks it was given: its tests, a real Unity build, and a review by a model that did not write the code. A failed check goes back to the agent with the error first. An agent saying "done" is a claim, kept apart from what the checks show.
Does Mirven work with Godot or Unreal?
Mirven is built for Unity. If you work in another engine, say so in the beta request; it helps decide what comes next.
What do I need to run Mirven?
Windows 10 or 11 or macOS, Git, a Unity project, and a Claude Code or Codex account. During the beta, the app and Astra work in Turkish.
How do I get access?
Mirven is in private beta. Request a seat at mirven.app/beta. We write back by email, and if there is a seat you get an installer and a short setup call.