Is there an AI copilot inside the Godot editor?

Ziva runs as an AI copilot dock inside the Godot editor in Godot 4.2 and later on Windows, macOS and Linux, since Godot itself ships no built-in AI assistant. Ziva's editor tools include get_scene_tree, get_godot_errors, get_screenshot, run_tests and execute_script, and every turn can be undone because Ziva snapshots project files before changes, a system rebuilt on Git in version 2.5.0. Ziva is free for hobby use with $3/mo included usage, or $20/mo on the Basic plan, while alternatives such as Claude Code or Cursor paired with a free MIT-licensed Godot MCP server reach the editor only through MCP tool calls.

Is there an AI copilot inside the Godot editor? Yes. Godot ships no AI assistant of its own its features page https://godotengine.org/features/ lists none , but Godot’s editor plugin system https://docs.godotengine.org/en/stable/tutorials/plugins/editor/making plugins.html lets one run inside the editor, and Ziva https://ziva.sh/docs does exactly that in Godot 4.2 and later on Windows, macOS and Linux. “Copilot” here means the generic thing: an assistant that works where you already work, sees what you see, and acts on it. Whether a tool qualifies comes down to one question. Does the AI run inside the editor with access to the scene tree, the errors and the running game, or does it read files from outside and reach the editor through a bridge? This post covers both kinds, what each costs, and what the in-editor option has shipped in measured tests. TL;DR | Option | Where the AI runs | Sees the live editor | Cost | |---|---|---|---| | Ziva | A dock inside Godot | Yes: scene tree, editor errors, running game | Hobby free with $3/mo included usage; Basic $20/mo | | Claude Code + a Godot MCP server | Terminal, IDE or desktop app | Through MCP tool calls | Claude Pro https://claude.com/pricing includes Claude Code: $20/mo billed monthly, $17/mo billed annually. The MIT servers Coding-Solo https://github.com/Coding-Solo/godot-mcp , hi-godot https://github.com/hi-godot/godot-ai are free | | Cursor + a Godot MCP server | VS Code-based editor | Through MCP tool calls | Cursor Pro https://cursor.com/pricing is $20/mo and is the cheapest plan that lists MCPs. Hobby is free | | A chat model in a browser tab | Browser | No | Free tiers exist Claude Free is $0 https://claude.com/pricing | What “inside the editor” means Ziva installs like any other Godot plugin https://ziva.sh/docs/installation and opens a dock. The agent in that dock understands the whole project https://ziva.sh/docs : scenes, nodes, scripts, resources, project settings, editor errors and the running game. When it needs something, it calls an editor tool rather than guessing. The Godot AI page https://ziva.sh/godot-ai names the tools: get scene tree to read the open scene, get godot errors to read what Godot logged, get screenshot to capture what it sees in a playtest, run tests to verify a change, and execute script to run GDScript for advanced changes changelog 2.5.0 https://ziva.sh/changelog . Two details separate this from a smart text editor. First, Godot errors sent to the agent include the source file and line 3.1.2 https://ziva.sh/changelog , and since 3.2.5 a tool result includes the script errors the operation caused, so the agent can see and address them. Second, every turn can be undone https://ziva.sh/docs . Ziva snapshots your files before it changes them, and 2.5.0 rebuilt that snapshot system on Git changelog https://ziva.sh/changelog . Undo restores the saved project files https://ziva.sh/ . What the copilot does in Godot The homepage https://ziva.sh/ keeps the maintained list. In short: - Builds and edits scenes, nodes, scripts, signals and resources. TileMapLayer tools require Godot 4.3 or later. - Writes and refactors GDScript and C with project context. - Searches the Godot documentation and reads editor output and errors. - Runs your project’s tests and can write tests to check its own work changelog https://ziva.sh/changelog , 2.4.0 . - Generates a sprite or texture through one generate image tool with RetroDiffusion and ChatGPT backends 3.2.1 https://ziva.sh/changelog , and adds a 3D model to the scene. - Launches a scene in the Game tab, sends keyboard and mouse input, and screenshots the result 3.0.0 https://ziva.sh/changelog . The separate Playtest Agent https://ziva.sh/full-game-tester plays through a whole game and is on Pro and Ultra. - Drafts a plan for approval in Plan mode before touching the project, and answers questions without changing anything in Ask mode slash commands https://ziva.sh/docs/slash-commands . - Reads an AGENTS.md https://ziva.sh/docs/agents-md for project conventions and takes @ file mentions https://ziva.sh/docs/mentions as context. What it costs The pricing table https://ziva.sh/ pricing has four self-serve plans. Hobby is free: $3 of included AI usage a month, 50 Ziva Baby messages a day, 10 Ziva Smart messages, no credit card. Basic is $20/mo and removes the message caps on Baby and Smart. Pro is $50/mo and adds the Genius lane, Sonnet and Opus models, and the Playtest Agent. Ultra is $200/mo. Two routes do not draw on your included Ziva usage. You can connect a Claude Code or ChatGPT Codex account https://ziva.sh/docs/existing-subscriptions on any plan, Hobby included, and requests count toward that provider’s usage. You can also run Ollama or LM Studio models locally https://ziva.sh/docs/local-providers . Separately, since 3.2.5 you can use supported models with your own OpenRouter credit changelog https://ziva.sh/changelog . Evidence it works We publish the runs, with costs, rather than adjectives. In the GPT-5.6 benchmark https://ziva.sh/blogs/gpt-5-6-benchmark-godot , each model got three prompts to build a playable Flappy Bird inside a live Godot editor through Ziva. GPT-5.6 Luna finished in 26 tool calls and 2.4 minutes for $0.17, Terra for $1.15, and Sol in 46 calls for $1.59. All three shipped with zero GDScript errors and scored 10/10 on the rubric. Meta’s Muse Spark 1.1 spent $3.05 and never finished the “playtest it and fix what you notice” turn, so it did not ship in the picker. The one-prompt parkour benchmark https://ziva.sh/blogs/godot-ai-benchmark is harder: build a 3D world with three races from a Kenney asset pack. GPT 5.6 Sol scored 6 for the world and 9 for the conversation at $2.94 in 9.7 minutes; Kimi 3 scored 8 and 7 at $3.14 but took 79 minutes; Claude Fable 5 scored 7 and 6 at $13 with caching $46 billed in 32.5 minutes. The GPT-5.6 post https://ziva.sh/blogs/gpt-5-6-benchmark-godot scored one run per model, two for Muse Spark. Treat these numbers as samples, not distributions. Outside benchmarks, Jody Mitoma asked Ziva to find the biggest memory, CPU and GPU costs in Beat Boxy’s levels. He reported the game ran “even smoother than it already was” and put the total cost at 58 cents Beat Boxy story https://ziva.sh/blogs/made-with-ziva beat-boxy . Railroad Rummy by Aterx Hobby, built with Ziva, is in Steam Early Access games made with Ziva https://ziva.sh/blogs/made-with-ziva . Copilots that live outside the editor Cursor https://cursor.com/ and Claude Code https://code.claude.com/docs/en/overview are general coding agents. Cursor is based on the VS Code codebase https://cursor.com/docs/configuration/migrations/vscode ; Claude Code runs in your terminal, IDE, desktop app and browser. On a Godot project both work from the files on disk. Godot’s TSCN scene format https://docs.godotengine.org/en/stable/engine details/file formats/tscn.html is text that the docs call “mostly human-readable”, and the same page covers .tres resources. Scenes and resources read as text alongside .gd scripts. Neither had a Godot page in its docs index as of September 18, 2026 Cursor https://cursor.com/docs/llms.txt , Claude Code https://code.claude.com/docs/llms.txt . They reach the live editor through an MCP server. Ziva ships one: Ziva as an MCP server https://ziva.sh/docs/ziva-mcp-server exposes the scene tree, node and script creation, error reading and run/stop to Claude Code, Codex, Cursor, VS Code, Windsurf, Zed and other clients with one-click setup, and it runs only on your machine. Open-source alternatives exist, headed by Coding-Solo/godot-mcp https://github.com/Coding-Solo/godot-mcp MIT, 5.7k stars, 14 tools for launching the editor, running the project and capturing debug output and hi-godot/godot-ai https://github.com/hi-godot/godot-ai MIT, 46 tools and 120+ operations, Godot 4.7+ . Be honest about the trade. Cursor’s Tab https://cursor.com/help/ai-features/tab is an AI autocomplete that suggests code as you type. Ziva is a chat agent and has no inline completion. If most of your day is typing GDScript by hand, Cursor’s completion is worth having, and the two tools can share one project through the MCP route. The Cursor comparison https://ziva.sh/blogs/ziva-vs-cursor-for-godot and the Claude Code comparison https://ziva.sh/blogs/ziva-vs-claude-code-for-godot go feature by feature. Related reading - How do I generate GDScript with AI? https://ziva.sh/blogs/generate-gdscript-with-ai : three routes, step by step - Is there an AI agent that can playtest a Godot game? https://ziva.sh/blogs/ai-agent-playtest-godot-game : the Playtest Agent in detail - Best AI Tools for Godot in 2026 https://ziva.sh/blogs/best-ai-tools-for-godot-2026 : the wider tool landscape - Ziva vs Godot MCP https://ziva.sh/blogs/ziva-vs-godot-mcp : the open-source bridge family compared