Ziva Release: The Best AI Godot Plugin Upgrades to Free GPT Mini
Mar 6, 2026
Two big changes in this release: Free tier users now get GPT Mini instead of our previous base model, and there’s an Undo button that lets you revert any file change Ziva makes with one click.
That alone would be a solid update. But we also shipped message queuing, in-editor docs search, a test runner, better scene tree filtering, and fixed a long list of bugs that were quietly making the plugin worse than it should have been.
If you’ve been looking for an AI Godot plugin that actually works reliably inside the editor, this is the update where we earned that. If you’re already a Ziva user, update your plugin. If you’re new, the free tier has never been this good.
New Features
Undo Button
This was our most requested feature. Before this update, if Ziva modified a file and you didn’t like the result, you had to manually diff and revert. Now Ziva takes a snapshot of every file before changing it. One click to undo.
It sounds simple, but it fundamentally changes how you work with an AI agent for Godot development. You can be more aggressive with your prompts — ask Ziva to try something risky, and if it doesn’t work, revert instantly. No git stash dance. No Ctrl+Z spam. Just click Undo.
Message Queuing
You can now type and send messages while Ziva is still generating a response. They queue up and get processed in order. Previously, you had to wait for the current response to finish before sending anything new.
This matters when you’re iterating fast: “make the player jump higher” followed immediately by “also add coyote time” — you don’t have to wait between thoughts anymore.
Run Tests Tool
The new run_tests tool discovers and runs your project’s test suite directly from the chat. If you’re using GdUnit4 or any test framework that follows Godot conventions, Ziva finds your tests, runs them, and reports results back in the conversation.
This closes the loop on the AI development workflow: write code, run tests, fix failures — all without leaving the chat panel.
Search Godot Docs
The search_godot_docs tool looks up official Godot documentation directly inside the editor. Instead of Ziva guessing at an API from training data (and sometimes getting it wrong), it can now look it up in real time.
This is a meaningful accuracy improvement. LLMs hallucinate API details — wrong method signatures, deprecated properties, parameters that don’t exist. Having direct access to the docs means Ziva’s answers are grounded in the actual Godot API, not a probabilistic guess from 2024 training data.
Thumbs Up / Down Voting
You can now rate assistant messages with a thumbs up or down. This goes directly into our quality metrics. We use it to find where models are failing on real Godot tasks so we can improve prompts, tool selection, and model routing.
Editor & Tool Improvements
Smarter Scene Tree Queries
The get_scene_tree tool now supports filtering by name, type, and group — with offset, limit, and depth controls. Before, querying the scene tree on a large project meant dumping hundreds of nodes and hoping the model could find what it needed in the noise.
Now you can ask for exactly what you want: “show me all CharacterBody2D nodes in the enemies group, 3 levels deep.” The AI gets a focused result instead of a 500-node dump, which means better tool calls, fewer hallucinations, and less wasted context window.
Better File Search
We replaced the old search-files and get-filesystem-tree tools with proper glob and grep implementations. The old tools were limited and didn’t handle edge cases well. The new ones work like you’d expect if you’ve used rg or find in a terminal — pattern matching, recursive search, and sensible defaults.
Editor Annoyance Fixes
Two small but important changes:
- “Files Modified on Disk” dialog is now auto-suppressed. When Ziva writes a file, Godot would pop up a modal asking if you want to reload. Every single time. Now it handles this automatically.
- “Reload scripts on external change” is enabled automatically. No more stale scripts in the editor after Ziva modifies them.
These were the kind of papercuts that made the plugin feel rough. They’re gone now.
Draft Recovery
If your message fails to send (network hiccup, server error), your chat input text is restored. You don’t lose what you typed. Small thing, but losing a carefully written prompt to a transient error is infuriating.
Error Visibility
Two changes here:
- Blank screen on load failure is fixed. If the plugin failed to initialize, it previously showed a blank white panel with no indication of what went wrong. Now it actually tells you.
- All tool cards are expandable. Previously, some tool call results were collapsed with no way to see errors. Now you can always expand them to see what happened — especially useful when a tool call fails.
Code Block Improvements
Higher contrast syntax highlighting that meets WCAG AA accessibility standards. JSON in tool outputs is now pretty-printed instead of crammed into a single line. These are readability improvements — you spend a lot of time reading code in the chat panel, and it should be easy on the eyes.
Bug Fixes & Stability
This is the section we’re most proud of. We fixed a lot of things that were silently broken.
Crash Fixes
save_scenecrash during rapid parallel tool calls. When Ziva made multiple changes at once (which it does often), concurrent scene saves could crash the editor. This was a race condition in how we coordinated with Godot’s scene system. Fixed with proper write locking.get_scene_treecrash on deleted script references. If a node referenced a script that no longer existed on disk, querying the scene tree would crash instead of handling the missing reference. Fixed.- Conversations permanently breaking from empty assistant messages. A malformed API response could insert an empty message into the conversation history, which then caused every subsequent load of that conversation to fail. These zombie conversations are now handled gracefully.
Windows Fixes
Windows users had a rough time before this update:
- Copy/paste fixed. WebView2 (the embedded browser we use for the chat UI on Windows) had broken clipboard integration.
- Installer extraction errors fixed. Some users couldn’t install at all due to archive extraction failures.
ripgrepnot found fixed. The bundledrgbinary wasn’t being found on some Windows configurations.- WebView2 focus stealing fixed. The chat panel would steal keyboard focus from the Godot editor. You’d be typing in the script editor and suddenly your keystrokes would go to the chat. This was maddening, and fixing it required digging deep into WebView2’s focus management APIs.
Silent Failures Fixed
edit_filesilently succeeding without writing. The tool would report success, but the file on disk was unchanged. Imagine asking Ziva to fix a bug, getting a “done” response, and nothing actually changed. This was one of the worst bugs we’ve shipped, and we’re glad it’s gone.- Plugin nagging for updates when GitHub is down. The update checker would fail to reach GitHub and interpret the failure as “new update available.” Fixed to only notify when we actually confirm a newer version exists.
- Parallel image generation timeouts. Generating multiple images concurrently would time out because we weren’t handling the async pipeline correctly.
- Concurrent SQLite write failures. Our local conversation storage uses SQLite, and under high concurrency (which happens when the AI makes rapid parallel tool calls), writes would fail silently. Fixed with proper write-ahead logging and retry logic.
- Dozens of silent fallback paths removed. This deserves its own paragraph. We found and removed a large number of code paths where errors were caught and silently swallowed — returning default values, empty arrays, or “success” statuses when something had actually failed. Every one of these was a debugging nightmare. If
get_scene_treereturns an empty array, is the scene actually empty, or did the query fail? You couldn’t tell. Now errors surface as errors. If something breaks, you’ll see it break.
Free Tier: GPT Mini
The free tier now uses GPT Mini. This is a significant upgrade in response quality for users who haven’t purchased a plan. Previously, free tier responses were noticeably worse — enough that it was hard to evaluate whether Ziva was actually useful without paying first.
That’s a bad experience. If you can’t tell whether a tool is worth paying for during the free trial, the trial is broken. GPT Mini is capable enough that you can build real things with it: generate scripts, create scenes, debug issues. It’s not our most powerful model (paid plans still get access to Claude Sonnet 4.5, Gemini, and others), but it’s good enough to show you what an AI agent for Godot development actually feels like when it works.
If you’ve been on the fence, now is a good time to try. We wrote a comparison of the best AI tools and plugins for Godot Game Engine in 2025 — the landscape has only gotten better since then, and we think this update puts Ziva firmly at the top for developers who want something that works inside the editor.
Update or Download
Existing users: The plugin should prompt you to update. If it doesn’t, re-download from ziva.sh/download .
New users: Install Ziva, open Godot, and start building. The free tier with GPT Mini gives you enough to evaluate whether this fits your workflow — no credit card required.