Wise choice.
OP opinion on that can be read here:
OP says:
“I hate that LLMs are laundering my GPL code into proprietary codebases, but I love that proprietary code bases are worse than ever (vibe coded slop) and that LLMs are also laundering proprietary code bases into open source.”
Hey folks, we’re gonna do a livestream at 2026-06-13T20:00:00Z on https://www.youtube.com/@danielginovker1984 to go over all of your questions. This includes the tough questions like LLM ethics in game development, all the points in @hkv and other folks’ comments, as well as the future of Ziva (we’re building a lot more cool stuff that will help with your Godot journey).
See y’all there! I’ll post a comment with timestamped responses as well afterwards for those of you who can’t make it.
Edit: Moved the Youtube channel to my personal one since YT requires 24 hour cooldown when applying to be a streamer or something
Livestream started ->> https://www.youtube.com/live/sVF12G2Ll0I
So I watched the livestream, and I have to say I was impressed with what I heard. You both seem to know what you are doing, are speaking intelligently about the subject and making a tool for creating games, not an AI slop machine.
From an ethical perspective, I think allowing people to use their own local models makes an excellent case for trying this out from an ecological impact point of view. It also would allow people in more countries, who may not have the financial resources to try it out.
I’m personally curious about what you do to prevent the models from hallucinating answers - as you mentioned that was something that Ziva prevented.
Appreciate the words dragonforge <3
I’m personally curious about what you do to prevent the models from hallucinating answers - as you mentioned that was something that Ziva prevented.
There’s a few things we do depending on the situation, but some quick examples
Editing C-Sharp → We dotnet build the modified script
Editing GDscript → We force static type hints then save and check for script/parse error
Editing a scene/node/tres → We reload through ResourceLoader, instantiate the scene, and validate every property against Godot’s real ClassDB
Edit TileMapLayer → We use the Godot API directly for set_cell, atlas setup, etc
Editing graphics → run the scene, grab the frame, hand it to a vision model to confirm it actually looks right
And then basically if any error happens, the edit is reverted, and the LLM is given an error message/relevant Godot docs it can use to understand why what it did was wrong.
Ziva v3 released!
This is our biggest release yet. So much we had to write a blog post to cover all the features. But some highlights:
This release also has something special: After working alongside developers with disabilities, we’ve added voice to text mode. You can now control most of the Godot editor with your own voice!
Grab it from the asset library at Ziva - AI Agent for Godot - Godot Asset Store
That’s a cool thing!
I don’t have any knowledge about how these work, could you very briefly explain how the A.I controls Godot.Does it just edits the .tscn ,.gd files as plain text? It’s quite interesting xD
Imagine a dev yelling at mic because the A.I is hallucinating xD
Jokes aside, it’s cool!
That’s really cool. Do you know if there is already a ticket open to adopt this in the core engine?
There’s 2 parts; the AI, and the Ziva plugin.
For the AI, it sees two things - your prompt, and a system prompt. The system prompt has “tools”, which are basically like “If you say the magic words edit_file("path/to/file.gd", "text_to_find", "text_to_replace"), then that text will get replaced”. Then the AI does it. Ziva has around 140 tools in total, everything from reading and writing files to diagnosing sprite sheet issues. Smarter LLMs are better at using the tools, but I spend a lot of time making sure even “dumb” LLMs can use these tools as well.
For the plugin, this is actually harder than it seems.. At first we would just edit files on disk, but if you have the same file open in editor, there’s a race condition about which one gets saved. This is further complicated if you’re actively editing the same file, since now there’s the unsaved file, the saved file, and the file on disk, and they can all be different.. This is just some really annoying reconciliation code lol
I don’t think it would fit in Godot core engine since it’s AI driven, but if anyone wants to collaborate, I might be down to work on some generic accessibility plugin for Godot 
That is a great invitation, of course, but personally I only dabble with Godot in my spare time. I might be able to find a small testing community via my employer, as accessibility is one of (many) our missions.
Ziva 3.0.1 and 3.0.2 are released ~
New Features
Also made a bunch of bug fixes for older Godot versions, older Linux CPUs and Macbooks, fixed some crashes, and improved compatibility with godot-mono on Linux
How opinionated is it in ‘how to code properly’? For instance I saw a few things in the video like “warning uncapped enemies could become unplayable”. But perhaps that was the whole point that there were eventually too many enemies to handle. Or perhaps it just meant the engine would grind to a halt (which is not good).
Also it suggested changes to make VSCode work better, but what if you don’t use VS code?
So these things and others made me think about its opinions on how stuff should be made, where it got them from and how opinionated it actually is.
I hope you don’t mind me asking, you have clearly put a ton of work into this and it is very impressive. I cannot imagine all the issues and problems you had to work around to achieve this.
I know that Copilot is quite opinionated, well, until you disagree and it immediately changes its mind. “This should be in a resource file … (tons of example code given)”. “I don’t want this in a resource file because…”. “You are absolutely right, this should not be in a resource file…”. Opinionated, easily swayed, completely sycophantic. I presume this still happens in your add on? (I don’t expect you to have ‘fixed’ AI in general of course) so that is why I am asking. Again, I hope you don’t mind me asking.
EDIT: You know I might have just answered my own question. I suppose it is as opinionated as whatever AI you point it towards using.
Ziva’s code quality is basically going to a combination of the LLM used, your prompt, and anything you put in AGENTS.md. In my personal projects I personally have extra rules about how scenes/etc should be managed, since otherwise LLMs tend to do their own thing
v3.0.3 and v3.0.4 are released!
Plus another 21 bug fixes, including fixing Linux resizing/scaling issues that have existed since day 1!
Read the full changelog: Changelog | Ziva
I’m mostly interested in how to use it with local models. My setup is a little bit different: on my workstation, I don’t run any AI harness or agent; instead, I run the model(s) on the GPU (AMD 16GB) with llama.cpp. My second PC (an old CPU with a mid-range Nvidia GPU) serves as my agentic system, which is remote-controlled via Sunshine and Moonlight.
I’m running two different setups. One uses two Gemma models (E4B and 12B in parallel), since the 12B can’t handle tool calls very well. The other uses a Qwen 3.6 27B running solo. IMHO Qwen is better at coding in Godot than Gemma, but it’s still not close to the frontier models. I’m running OpenWebUI there as a chat server.
The remote second system accesses it over LAN. On that system, I’m running OpenCode and VSCodium with Continue.
What local models have you tested with the Ziva harness? What local model would you recommend for Ziva? I haven’t tested Ziva yet, but I will do so in the future as it sounds interesting. I also watched parts of your livestream.
Honestly I don’t run models locally, but I asked in the Discord and consensus seems to be Qwen3-Coder-Next-UD_Q8_K_XL (coding) / Qwen3.6-27B-MTP (planning)
I’ll post update notes soon about all the releases that have gone out, but wanted to share a cool new feature we’re working on - Fully AI driven QA Playtesting. Here’s Ziva playing a Minecraft clone, chopping a tree and making a crafting bench out of it:
Much has changed since we last shared updates, so gonna do that all here!
We are currently on v3.1.5. Main content added over the last couple weeks:
We’ve also fixed like 100+ bugs. Check out the kind of output our full game playtester produces:
The day this gets added to Godot is the day I stop using it.