Introduction

Aura, the Unreal and Unity coding assistant from Ramen, has open-sourced Autopilot, an agent that plays your game inside the editor in real time and reports what breaks. The Apache-2.0 repo went up on October 30, the first v1.0.0 release shipped prebuilt plugins for Unreal Engine 5.4 to 5.8, Unity 6 and Godot 4.7+, and Aura

at about 23:32 MYT the same day, calling it "Computer Use for games."

You give it one sentence, like "Kill the boss while keeping your health above 20%," and it opens a play-in-editor window, reads your project's source, probes the running game, plays with real inputs and closes the window with a written report.

What shipped

  • Tool: Aura Autopilot, a real-time playtesting agent (product page, docs)
  • Builder: Ramen, the team behind the Aura editor agent (Unreal plugin binaries are signed by TribeVR Inc)
  • Modality: game-playing agent with vision, driving live keyboard, mouse and controller input in the editor
  • Open vs API: open source (Apache-2.0), but it calls hosted models. Bring your own keys, your Claude Code or Codex subscription, or an Aura plan
  • Where to run it today: as a plugin in Unreal, Unity or Godot on Windows or macOS, or through Claude Code and Codex via its skills and MCP server
  • Concrete fact: in Aura's seven-task Unreal Engine 5.8 benchmark, Autopilot passed 7 of 7 with a mean of 262 seconds per task, against 5 of 7 and 437 seconds for Claude Code

How it plays

The repo's .env.example shows the split. A language model (Claude Opus by default, through the OpenCode agent, or Claude Code or Codex if you switch AUTOPILOT_AGENT) researches your project and writes a harness for the game. A separate decision model called Jev, from Typesafe, does the actual playing. Model-written code runs in a sandbox, and Autopilot keeps a per-project memory of what it learned, so the next run starts smarter. That's the "self-improving" part of the pitch.

Prompts fall into four types, according to the docs:

  • Exploration: "Explore the crafting system in Lvl_Main."
  • Testing: "Go through the currently open level and tell me everything that looks wrong."
  • Verification: "Verify that the tree shakes when you hit it with an axe."
  • Goal: "Kill the boss while keeping your health above 20% of its maximum value."

Aura says it has run it on shooters, souls-likes, mobile match-3, survival crafting, turn-based JRPGs and real-time strategy games. Blueprint-only Unreal projects work too.

The benchmark claim

Aura's whitepaper ran Autopilot, Claude Code and an unnamed third game agent through the same seven tasks in one Unreal Engine 5.8 game. Every run started with empty memory, and any run that cheated counted as a failure. The product page reports:

AgentTasks passedMean time per task

Autopilot

7 of 7

262 s

Claude Code

5 of 7

437 s

Another game agent

4 of 7

508 s

That's the vendor's own test on one game, so treat it as a starting point rather than a verdict. The repo ships a bench/ folder with a runner and an example task file if you want to try your own.

Run it

Grab the zip for your engine from the latest release, then:

  • Unreal (5.4 to 5.8): copy Autopilot/ into your project's Plugins/ folder and click the orange Autopilot icon
  • Unity 6: Package Manager, then "Install package from tarball…" and pick com.ramenvr.autopilot-<version>.tgz, then open Window > Autopilot
  • Godot 4.7+: move autopilot/ into addons/ and enable it in Project Settings > Plugins

The setup flow nudges you toward an Aura account, but the open-source build also takes your own API keys. To drive it from a coding agent instead:

autopilot connect claude
autopilot connect codex

Then ask Claude Code or Codex to "Use Autopilot to verify the double jump works on every ledge in Level_02."

The honest notes

Aura is upfront about the limits. Replying to developers on X, the team said Autopilot catches anything that is a real bug and has some sense of "this looks wrong," but it is "really bad at taste," so judging fun stays with humans. It runs inside the editor for now, not on shipped builds, and mobile builds are something they're still exploring. Unity packages aren't signed yet. And since the planner and player are hosted models, each test run has an API or subscription cost.

Why it matters

Generative tools have made building a level or a mechanic much faster, but someone still has to play it to know it works. Autopilot is the first open-source agent we've seen that plugs into all three major engines and closes that loop from a one-line prompt. With Aura building a feature and Autopilot playing it, the build-and-test cycle starts to run on its own, and anyone can now read, fork or extend the tester.

—Titus