Ramen open-sources Aura Autopilot for in-editor game testing

Ramen has open-sourced Aura Autopilot, an in-editor agent the company built to playtest a game from a prompt and carry what it learned into the next run. A dispatch dated October 31, 2026, by Nyx Iskandar and Andy Tsen, says Autopilot is compatible with Unreal Engine 5, Unity 6, and Godot 4 on Windows and macOS. The repository is public at github.com/ramenvr/autopilot.

GamesBeat, covering the announcement, reports that the agent works with existing Claude Code and Codex subscriptions, that the standalone open-source version can use a developer's own API key, and that Autopilot also integrates with Aura subscriptions. Andy Tsen, Ramen's CEO and co-founder, said in the release that Autopilot "plays any game in real time and learns as it goes," and that open-sourcing it is meant to let developers and researchers take the work further.

Godot belongs to that Autopilot compatibility line, not to every Aura tool. GamesBeat still describes Aura as Ramen's assistant for Unity and Unreal, and says Aura 1.0's ongoing engine support is those two. The vendor site headers Aura the same way.

What the dispatch claims

Iskandar and Tsen describe Autopilot as a system of large language models and decision models, aimed at generality across games, engines, operating systems, and playtesting tasks. They say it has played, through real user-input events, real-time strategy, first-person shooters, survival crafting, turn-based role-playing, match-3, and Soulslike games.

Four components do the work. An engine plugin reads game state and injects real user inputs. A Python agent service hosts the coding agent, the decision-model runtime, a sandbox, and per-game memory, and talks to the plugin over a WebSocket bridge. The coding agent researches the project and writes a harness; Ramen says that agent can use the SDKs for Claude Code, Codex, or OpenCode. The decision model, which the dispatch says is Jev via the TypeSafe SDK, chooses actions during play.

The harness is written per game, not shipped as one universal script. Ramen's account is that, because Autopilot sits in the editor, the coding agent can read source and play-in-editor state and emit macro-actions on a shared parent class, rather than handing the decision model a raw list of keypresses. Continual learning here is a per-project memory filesystem: a game codex, fact notes, reusable harness code, and traces from earlier runs. Later runs can skip research already done. Ramen says the memory still flags changed files so a later harness can account for source edits.

Company numbers, not an outside score

GamesBeat's speed and accuracy lines come from a press release, not an independent test. Those highlights say Autopilot runs 3x faster than Aura's Verification Agent at a lower cost, and that it beats Bezi, NeoStack, Claude Code, and Codex at testing games. That is Ramen's claim, repeated by GamesBeat.

The dispatch reports a different company-run comparison, on an unmodified Easy Survival RPG V5 project in Unreal Engine 5.8. Memory was emptied before each run. A pass required the expected verification verdict, or the stated goal, within 1800 seconds, without cheating such as teleporting the player or spawning items. Ramen reports Autopilot at 7 of 7, Claude Code at 5 of 7, and a leading Unreal/Unity agent it labels only as Competitor at 4 of 7. Mean time, from first generated token to last and excluding play-in-editor launch, was 262 seconds, 437 seconds, and 508 seconds. The runs used Claude Opus 5.5 on medium effort. That table does not name Bezi or NeoStack, and it is not the 3x comparison against Aura's Verification Agent.

Ramen has also published a white paper on the architecture and its benchmarks.

Read the dispatch before you treat any of those figures as settled, then open github.com/ramenvr/autopilot and check a cold play-in-editor prompt on a project whose correct outcome you already know.