Autopilot plays and tests your video games for you

Anyone who has ever built a video game knows the pure torture of manual playtesting. Running through the exact same dungeon thirty times just to catch a single clipping glitch or a broken quest trigger is exhausting, mind-numbing, and prone to human error.

Aura has open-sourced Autopilot under the Apache 2.0 license, a real-time autonomous playtesting agent designed to play through your builds, learn the rules, and hunt down bugs run after run.

Drop in the plugin, prompt an objective, and watch it play

Autopilot integrates directly into your engine as a native plugin across Windows and macOS. It currently supports Unreal Engine 5 (versions 5.4 through 5.8, with source builds also supported on Windows), Unity 6 (v6000.0 or later), and Godot 4 (v4.7 or later).

Setting it up avoids the headache of writing brittle test scripts. The workflow is driven by natural language prompts across four core buckets: exploration, general testing, feature verification, and specific goals. You can tell it to explore the crafting system in a specific level, verify that trees shake when chopped with an axe, or beat a boss while staying above 20% health.

Beyond clicking the orange controller icon inside the editor, Autopilot also exposes an MCP server. With a quick terminal command, developers can hook it into Claude Code or Codex and trigger playtests straight from their coding assistant.

Reads your source, builds harnesses, and visibly gets better

Once prompted, Autopilot launches a Play In Editor (PIE) window. Instead of treating the game purely as black-box video frames, an LLM coding agent reads your project source code, inspects the active level, and writes dynamic test harnesses for an underlying decision model.

During live gameplay, that decision model takes the wheel. Early runs can look clumsy, but the system continuously populates a persistent project memory. You can even spin up a ten-minute autonomous exploration run right after onboarding to let it warm up. The more Autopilot plays your game, the sharper it gets, demonstrating faster dodges, smarter weapon choices, and better aiming.

When the run wraps up, the game window closes automatically, and Autopilot spits out a clean diagnostic report detailing clipping issues, broken sequences, and unexpected behavior. Account-wise, developers can connect an Aura account or bring their own keys; running through MCP offloads LLM token costs to your existing Claude or ChatGPT subscription.

Why it matters

For solo indie developers and small studios, Autopilot looks like an immediate win. Dedicated QA teams are a luxury most small budgets cannot afford, and testing your own levels over and over breeds massive blind spots. Having an AI agent that understands your source code and repeatedly stress-tests combat and traversal overnight removes a brutal amount of grunt work.

Across the industry, this marks a clear shift away from brittle, hardcoded QA automation scripts that break whenever a designer tweaks an asset position. Pairing an LLM that parses code with a decision model that handles real-time gameplay is a remarkably clever architecture, and it will likely force established game-testing toolkits to pivot toward agentic workflows.

That said, we see clear boundaries to what it can do right now. Autopilot is fantastic for verifying logic, catching collisions, and validating mechanical balance, but an agent cannot tell you whether a combat loop actually feels satisfying or if a narrative beat lands emotionally. It eliminates the mechanical grind, but judging the fun still belongs entirely to humans.