Quick picks
| If you need… | Use |
|---|
| An open-source playtester inside UE5, Unity 6 or Godot 4 | Aura Autopilot |
| Agents that play your PC or mobile build on real devices | nunu.ai |
| Zero-integration bug hunting with video bug reports, for a team with a QA budget | modl.ai |
| Playtesting inside the Unreal or Unity editor while you build | Bezi (can start Play In Editor and drive inputs) |
| A coding agent that playtests your Roblox place | Claude Code, Cursor or OpenAI Codex through the Roblox Studio MCP server |
| Automated tests for gameplay code | Claude Code or Cursor writing Unity Test Framework, GUT or gdUnit4 tests |
Aura Autopilot: open-source in-editor playtesting (UE5, Unity 6, Godot 4)
Aura Autopilot is an open-source autonomous playtester from Ramen (the makers of Aura). Per Ramen's docs and the ramenvr/autopilot README (Apache License 2.0), it learns your game, explores levels and improves its gameplay run after run. Prebuilt downloads cover Unreal Engine 5 (5.4–5.8), Unity 6 (v6000.0+) and Godot 4 (v4.7+) on Windows and macOS; advanced Unreal users on Windows can also build from source against a custom engine. The product page frames it as an open-source game tester that plays in real time.
Prompt types (from the Autopilot 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") and Goal ("Kill the boss while keeping your health above 20% of its maximum value"). For each prompt Autopilot launches a PIE window, reads project source, plays according to the prompt and generates a report when the PIE window closes.
Three ways to run it:
- Engine plugin: download from tryaura.dev/autopilot, drop the plugin into Unreal's
Plugins/, Unity's Package Manager (tarball) or Godot's addons/, then complete onboarding in the Autopilot window. - Through Aura: if you already use Aura, Ramen says no extra Autopilot setup is needed beyond the Aura setup; Autopilot is also included in an Aura subscription. Aura's free trial (no credit card) includes starter Auto Mode usage and $10 of premium credit.
- Claude Code or Codex via MCP: after plugin install, run
autopilot connect claude or autopilot connect codex. With an Aura sign-in, Aura only counts decision-model usage and LLM tokens come from your Claude or ChatGPT subscription; BYOK needs a TypeSafe API key.
Self-improving claim (Ramen's, not ours): Ramen writes that Autopilot "visibly improves the more it plays your game" as it fills project memory (faster harnesses, smarter combat and aiming). Treat that as the vendor's claim until you verify it on your own title. Their Oct 31, 2026 update also says Autopilot has been evaluated on RTS, FPS, survival crafting, turn-based RPG, match-3 and Soulslike play with real input events.
What we haven't tested: we have not installed Autopilot in UE5, Unity 6 or Godot 4, run a prompt, or reproduced Ramen's whitepaper benchmarks. This section is sourced from the GitHub README, Autopilot docs, the Autopilot download page and Ramen's Oct 31, 2026 update only.
Sources: GitHub ramenvr/autopilot · Autopilot docs · Download page · Oct 31, 2026 update. Tool page: Aura Autopilot.
Godot: let Claude Code or Cursor playtest a scene (Ziva run_scene)
New on November 7, 2026: Ziva, a commercial AI agent plugin for the Godot editor, documented a run_scene tool that its local MCP server also gives to Claude Code, Cursor or Codex. One call runs a scene in the open editor, sends timed input (input actions, keys with modifiers, clicks at a position or on a button found by its text, mouse motion) and returns a screenshot of the final frame plus the run's console output and runtime errors, with a report of the frame each input landed on. That closes the fix, run, look loop for Godot the way Roblox Studio's MCP already does.
- Price: works on Ziva's free Hobby plan, but you must be signed in (screenshots are uploaded to Ziva). The goal-driven Full Game Playtest Agent, which returns pass, fail or blocked, is Pro ($50/month) and Ultra only and isn't exposed over MCP.
- Limits: the Godot editor must be open; no headless or CI mode, gamepad or touch input; one frame per call, not video.
- Ziva's own test: Claude Code plus
run_scene on a sample game found the planted score bug (a missing node and a null-instance error at main.gd:15) in an 8.1-second call. That's the vendor's run; we haven't installed Ziva.
For CI, script it instead. The same Ziva guide ran the free test frameworks headless on Godot 4.7.2 (November 5 and 7): GUT 9.7.1's GutInputSender and GdUnit4 6.2.3's scene runner both simulated the key press and caught the bug. GUT exited with code 1 on the failing test; GdUnit4 refuses --headless (exit 103) unless you add --ignoreHeadlessMode, then exited 100 on the failure. Also from that run: time GdUnit4 waits in milliseconds (await_millis), not frames, or the test can pass headless and fail under xvfb. These are useful prompts to hand your coding agent when it writes Godot tests. Setup for Ziva and the free MCP alternatives: Godot MCP setup.
Ziva sources (checked November 9, 2026): the Ziva as an MCP server docs; Ziva's Godot MCP server for Claude Code and Cursor (November 7, 2026) and How to automate playtesting in Godot (updated November 7); the changelog (3.3.6, November 5-7) and GitHub releases (v3.3.6, November 7); plans from ziva.sh and its Terms of Service (last updated November 7, 2026).
AI agents that play your build
These tools test your game from the outside, the way a human tester would: they look at the screen and press buttons. You describe the test in plain language, like “complete the tutorial” or “open the shop and buy the cheapest item”, and get a pass or fail plus a recording. Autopilot (above) is the in-editor cousin; nunu.ai and modl.ai work on packaged builds and devices.
nunu.ai runs its agents on PC and real iOS and Android devices (console support is listed as coming). Each step, the agent takes a screenshot, plans, and acts. An optional SDK lets you expose helper functions such as cheats, level unlocks or navigation to an objective, which saves the agent from slow trial and error. Self-serve plans include MCP and API access. Pricing from nunu.ai/pricing (October 2026): a free Starter plan with 2,000 one-time credits, then $200/month for 20,000 credits (nunu estimates about 40 hours of agent work) on one guaranteed device, picked from 60+ real iOS and Android devices.
modl.ai is fully black-box: upload a build, define tasks, and its agents use vision models and OCR to read the UI and play. Its analysts then report crashes, broken menus, missing or placeholder assets, softlocks and performance drops with video, logs and severity. Runs can be triggered from CI. modl.ai says it supports Android and desktop today with iOS in development, and works best on games with clear UI and structured interactions (match, narrative, card and turn-based games); very fast or timing-critical gameplay isn't fully supported yet. There's no public price list, so you book a demo.
Playtesting from inside the engine
- Unreal, Unity and Godot: Aura Autopilot (section above) installs as an engine plugin and playtests from plain-English prompts, including through Claude Code or Codex over MCP.
- Unreal and Unity: Bezi can start Play In Editor and drive inputs to test what it just built. Epic's official Claude Code plugin also exposes Unreal's automation-test tools over MCP. Setup steps are in our Unreal guide.
- Godot: Ziva's
run_scene runs a scene with timed input and returns a screenshot and errors, inside Ziva or through Claude Code or Cursor over its MCP server (section above). - Roblox: with Studio's built-in MCP server turned on, a connected agent can start and stop playtests, read the output log and take viewport screenshots, so it can fix a bug and check the fix in one loop. It can also send keyboard and mouse input and hand a test to Roblox's playtest subagent. Setup and limits: Roblox Studio MCP setup; the wider toolkit: AI tools for Roblox.
- Unity: Unity AI's Assistant can write C# and run automation scripts in the Editor. See AI tools for Unity.
- Unity, Unreal or Godot from one app: AI Game Developer can playtest through the engine plugins it installs (vendor claim; untested by us). See AI Game Developer.
- Web games: the open-source OhMyGame desktop app (beta) opens the browser game its agent built in a window and plays it, including canvas and WebGL games, then tries a fix and plays again. Claude Code or Codex can borrow that playtesting through its MCP server.
Let AI write your automated tests
The cheapest QA win for a solo dev is test coverage you'd never write by hand. Ask a coding agent to write tests for the logic that breaks most often: save and load, inventory math, damage formulas, quest state, procedural generation seeds.
- Pick the framework: Unity Test Framework for Unity, GUT or gdUnit4 for Godot.
- Tell the agent your engine version and point it at one system at a time.
- Have it run the tests and fix failures, and review each fix so it doesn't “fix” a test by weakening it.
- Run the suite on every commit (for example with GitHub Actions), so regressions show up the day you cause them.
AI for human playtest feedback
AI doesn't replace real players, because only people can tell you whether the game is fun. It does help you process their feedback: paste survey answers, Discord threads or Steam playtest comments into an AI chat and ask it to group them into themes, count how often each comes up, and pull out exact repro steps for bugs. Remove players' personal details first.
Shipping on Steam? Steam Playtest gives you a free, separate app ID for outside testers whose reviews and playtime don't touch your main game; our Steam launch checklist shows where it fits before a demo and Next Fest, and Steam Playtest vs demo vs Early Access explains when a Playtest beats a demo or paid Early Access.
Which should you use?
- Solo or hobby UE/Unity/Godot dev: try Aura Autopilot's free download (Aura trial or BYOK) plus AI-written unit tests.
- Mobile game near launch: try nunu.ai's free credits on your menu, onboarding and purchase flows across devices.
- Studio with a QA budget: compare nunu.ai and modl.ai on the same build and test list.
Also see: how to vibe code a game · best AI tools for solo game devs · AI tools for Godot · put a web game on Steam (includes Steam Deck checks for wrapped builds) · controller support for browser games (a hot-plug and multi-browser test list).
Get the Starter Stack ($12) Get the free weekly digest