Yes. Ziva’s Full Game Playtest Agent shipped in plugin 3.1.4 in August 2026. It launches your Godot game on your machine, sends real keyboard and mouse input, and ends with a pass or fail verdict plus a saved video. It needs the Pro plan.
There are two levels of this. The distinction matters for what you pay. Turn-by-turn playtesting has been part of Ziva since 3.0.0. The agent launches a scene, sends input and screenshots the result. It is used inside ordinary chats. The GPT-5.6 benchmark shows one turn’s tool calls: two files, a scene-tree read, physics validation, and a self-playtest. The Full Game Playtest Agent is a separate subagent. It plays a whole game against a goal you set. This post covers both levels, what the benchmarks showed, and what they cannot do yet.
TL;DR #
| Level | Plan | Since | What you get |
|---|---|---|---|
| Turn-by-turn playtest | Core feature | 3.0.0 | Agent launches a scene in the Game tab, sends input, screenshots the result |
| Full Game Playtest Agent | Pro and Ultra | 3.1.4. Runtime errors since 3.1.5. Recordings since 3.1.6 | Plays the whole game to a goal, saves a video, returns a verdict |
How the Playtest Agent works #
The Full Game Tester page describes the loop in four steps:
- It reads the code first. The agent analyses the game to work out how to reach the goal and where to add temporary hooks and logging.
- It launches the game and freezes it. LLMs do not work in real time, so the game runs in a sandbox and is d while the model thinks.
- It plays and measures every move. Real inputs go to the game. Output is measured through screenshots, custom hooks the model can query, and the added logging.
- It ends with a verdict. The model reviews the whole run for pass or fail and saves a video for you.
You give it a test to run, such as “the score goes up when the player clears a pipe”, and it plays until it sees that happen or gives up. Since 3.1.5 runtime errors from the running game are fed to the agent and its thinking is visible in the chat. Since 3.1.6 it records the session with its goal and the inputs it is sending captioned on screen, and it is told what happened in the game rather than only what its own probes saw.
Two real runs are on the page. In Godotcraft the agent is told to chop down a tree and make a crafting bench. It identifies movement, pathfinding, tree blocks and crafting. It walks to a tree, chops a log, and builds the bench. The video is graded pass or fail. In a mini-golf game it lines up the shot while the game is d. It holds ui_left and finishes the hole in two strokes.
What playtesting changed in our benchmarks #
The benchmarks use the turn-by-turn tool. How a model uses its playtests shows up in all three posts.
In the GPT-5.6 Flappy Bird benchmark, prompt two was “Add a score counter that ticks when the bird passes a pipe, then playtest it and fix what you notice”. We verified each game with three playtests of our own. We let the bird die, flapped through a pipe, and triggered the restart. GPT-5.6 Luna, Terra and Sol all shipped at 10/10 on the rubric for $0.17, $1.15 and $1.59. Muse Spark 1.1 turned the same prompt into an edit-playtest-edit spiral. It hit the 20-minute cap on all three turns and did not ship.
In the parkour benchmark, Kimi 3 ran 12 playtests and produced the best world (8/10). It took 79 minutes. GPT 5.6 Sol playtested only where it counted. It finished in 9.7 minutes for $2.94. Claude Fable 5 had the sharpest debugging. It never saw its own playtest frames and scored 6/10 on the conversation.
The vampire-survivor benchmark is the cautionary tale. GPT 5.6 Terra ran 6 playtests and reacted to none of them. That is how it shipped attacks that never connect. Claude Opus 5 looked at 11 of its 16 playtests. It caught a 1 FPS collapse from a single frame. It scored 9/10 on the conversation at $65.64. Playtesting only helps a model that reads the result.
Limits to know #
- Engine. Godot 4.2 and later today. Other engines go through the enterprise program (Full Game Tester page ).
- Plan. The Playtest Agent needs Pro at $50/mo. Turn-by-turn playtesting in ordinary chats is a core feature, and a setting turns the playtest tool off entirely if you never want the agent to run the game (3.1.2).
- Input reach. Our harness verified restart buttons in code and through keyboard paths because it could not click CanvasLayer buttons at the time of the GPT-5.6 benchmark . That note is from July and covers the turn-by-turn tool. We have not published a re-test.
- Cost. A full playtest is many model turns with screenshots, so it costs more than a chat message. The benchmark posts above give the cost of whole builds that used the turn-by-turn tool. We have not published per-run costs for the Playtest Agent.
- Data. Ziva does not store your prompts, code or project data from AI requests on its servers. It never trains on your code. A cloud model processes the context sent for the task under that provider’s own retention terms (privacy policy ). The Playtest Agent always calls a hosted model for thinking and image analysis. The game itself runs locally and your project files stay on your machine (Full Game Tester page ). In ordinary chats you can pick a local model throughOllama or LM Studio .
- What it measures. A pass or fail against a goal. It does not tell you whether the game is fun.
Alternatives #
| Tool | Drives game input | Verdict and video | Licence and price |
|---|---|---|---|
| Ziva Playtest Agent | Yes, keyboard and mouse | Yes | Pro plan, $50/mo ( [pricing](https://ziva.sh/#pricing) ) |
| Ziva turn-by-turn playtest | Yes | Screenshots only | Core feature ( [3.0.0](https://ziva.sh/changelog) ) |
| [GDAI MCP](https://gdaimcp.com/) | Input simulation and screenshots | Not listed. Any verdict comes from the client you attach | $19 one-time, [binary-only licence](https://gdaimcp.com/license) |
| [Coding-Solo/godot-mcp](https://github.com/Coding-Solo/godot-mcp) | No input tool listed. Has `run_project` and`get_debug_output` | Not listed | MIT, free |
| GUT | Scripted only, via GutInputSender . No agent | Pass or fail from asserts you write. Video not listed | MIT , free |
GDAI MCP and the open-source servers hand the loop to whichever client you attach, so the verdict depends on that client. Our May comparison with GDAI MCP predates live playtest (3.0.0) and the Playtest Agent. It said Ziva could not drive input and described GDAI as free and open source. GDAI’s licence page covers the compiled binary only, and its homepage lists $19 one-time. The table above supersedes that post. GUT is the right tool for logic you can assert in code. Its GutInputSender replays inputs you script in a test. It does not decide what to press. You write the asserts that judge the result.
Related reading #
- [Is there an AI copilot inside the Godot editor?](https://ziva.sh/blogs/ai-copilot-inside-godot-editor) : the in-editor agent that runs these playtests
- [How do I generate GDScript with AI?](https://ziva.sh/blogs/generate-gdscript-with-ai) : the write-then-verify loop for scripts
- [Godot Benchmark: Sol > K3 > Fable](https://ziva.sh/blogs/godot-ai-benchmark) : 12 playtests versus none seen
- [GPT-5.6 Benchmark: We Made It Build a Godot Game](https://ziva.sh/blogs/gpt-5-6-benchmark-godot) : the “playtest it and fix” prompt