GODOT LLM
Godot LLM: use the model you want inside the editor
Ziva connects the LLM you choose directly to Godot, so it can write GDScript, fix errors, and edit your real project. Start with Ziva's free LLM, choose from 14 hosted models including Claude, GPT, Gemini, Grok, Kimi, and DeepSeek, or run a local model through Ollama or LM Studio.
Free LLM included - no API key required
Leading LLM providers
Use models from the top AI companies in Godot
Anthropic
Claude
OpenAI
Codex
Gemini + Antigravity
DeepSeek
DeepSeek
MiniMax
MiniMax
Moonshot AI
Kimi
xAI
Grok
The lineup
Every LLM you can point at your Godot project
One account, one dropdown, no per-provider setup. Switch model between messages - drafting with a cheap one and finishing with a frontier one is the normal way to work.
| Model | Lab | Context | Your prompts | Plan |
|---|---|---|---|---|
| Ziva Free | Ziva | 200K | Up to 30 days, no training | Free |
| Claude Fable 5 | Anthropic | 1M | Up to 30 days, no training | Pro |
| Claude Haiku 4.5 | Anthropic | 200K | Zero retention | All plans |
| Claude Opus 5 | Anthropic | 1M | Zero retention | Pro |
| Claude Sonnet 5 | Anthropic | 200K | Zero retention | Pro |
| Deepseek v4 Pro | DeepSeek | 200K | Up to 30 days, no training | All plans |
| Gemini 3.7 Flash | 200K | Up to 30 days, no training | All plans | |
| Minimax M3 | MiniMax | 1M | Not published | All plans |
| Kimi K3 | Moonshot AI | 1M | Up to 30 days, no training | All plans |
| Kimi K3 Turbo | Moonshot AI | 128K | Up to 30 days, no training | All plans |
| GPT 5.6 Luna | OpenAI | 1M | Up to 30 days, no training | All plans |
| GPT 5.6 Sol | OpenAI | 1M | Up to 30 days, no training | Pro |
| GPT 5.6 Terra | OpenAI | 1M | Up to 30 days, no training | All plans |
| Grok 4.6 | xAI | 500K | Up to 30 days, no training | All plans |
“Your prompts” is what the lab does with the code Ziva sends it. Zero-retention models keep nothing; the rest hold prompts up to 30 days for abuse monitoring without training on them, and where a lab publishes no policy we say so rather than guess. Ziva stores no project data either way. The free model costs nothing at any tier, and the Hobby plan includes $3 of monthly usage on the rest.
Context, not copy-paste
A Godot LLM that understands your project
In Ziva, your LLM can inspect the project before it answers. That means fewer guessed node paths, fewer incomplete snippets, and changes that fit the game you are building.
Search your whole Godot project
Ask a question and the LLM finds the relevant scenes, scripts, signals, and node names itself. You do not have to collect files or paste code into a browser chat.
Understand your live scene tree
Ziva gives the model the structure of your open scene, so it works with the nodes and paths that are actually in your game instead of guessing.
Fix real Godot errors
The LLM can read editor and runtime errors directly, trace them back to your code, and fix the problem without asking you to describe the output.
Make changes you can review and undo
The model edits your Godot files in place and shows you what changed. You can review every edit and undo it like any other project change.
These are four of the tools the agent can call. See the full Godot AI assistant for the whole surface, including scene editing, tilemaps, and art generation.
Bring your own LLM
Use a local LLM or a subscription you already have
Run local models through Ollama or LM Studio for free, with your code and prompts kept on your machine. Or connect Claude Code and Codex over OAuth instead of paying for the same model twice.
Connect Claude Code or Codex with OAuth
Sign in with your Claude Code or ChatGPT Codex account and Ziva uses the subscription you already pay for. Your account and usage limits stay yours - on every Ziva plan, including the free one.
Or point your own LLM client at the editor
Ziva exposes the whole Godot editor as a Model Context Protocol server. Any MCP-aware client - Claude Code, Cursor, or your own agent - gets the same typed tools as the in-editor model, so it can search and edit a large Godot project in place.
Which LLM should you use?
We gave three LLMs the same Godot prompt
One prompt - build a 3D parkour world with three timed races against bots - run unattended through Ziva, scored on the world it produced, how it held a conversation, what it cost, and how long it took.
| Model | World | Conversation | Cost | Time |
|---|---|---|---|---|
| GPT 5.6 Sol | 6/10 | 9/10 | $2.94 | 9.7 min |
| Kimi K3 | 8/10 | 7/10 | $3.14 | 79 min |
| Claude Fable 5 | 7/10 | 6/10 | $13 | 32.5 min |
Fable's $13 is what the run cost with prompt caching; at list price the same run bills $46.
Kimi built the better world and took eight times as long to do it; Sol was good enough, fast, and cheap. There is no LLM that wins every task, which is the argument for keeping all of them one dropdown away. Read the full benchmark, with the worlds you can play.