Orca: Open-Source AI Agent for Godot

AI agent that feels at home in the Godot editor.

Orca is an open-source, MIT-licensed AI development agent built directly into Godot. It brings project exploration, editor awareness, Godot-specific tools, reviewed changes, and focused runtime workflows together in one integrated editor dock.

Orca is more than a chat interface beside your project. It understands the active scene, selected nodes, open script, caret position, project settings, saved resources, and other relevant editor context. It can investigate a task, explain what it finds, maintain a visible plan, and prepare precise changes without silently taking control of your files.

Use a supported cloud provider or connect a model running locally through Ollama, LM Studio, or another OpenAI-compatible server. Whichever provider you choose, Orca keeps its work visible and leaves project changes under your control.

Move from read-only planning to reviewed implementation without losing control of the workflow.

Why Orca

Built into Godot

Orca runs as a Godot EditorPlugin, not as a separate application or browser window. Its interface follows the editor's scale and workflow, and relevant scripts, scenes, resources, project settings, and Godot Help topics open in the editor where they belong.

Designed for Godot projects

Orca works with Godot concepts instead of treating a project as an unstructured collection of text files. It can:

• Inspect saved scene hierarchies, serialized properties, groups, child instances, and signal connections.
• Read focused GDScript functions, including the exact unsaved source currently open in the script editor.
• Inspect selected project settings, Input Map actions, serialized resource dependencies, diagnostics, and reflected Godot API signatures.
• Propose reviewed changes to scripts and text files, Input Map actions, the main scene, selected display settings, and constrained scene structures.
• Open relevant files and Godot documentation directly from its activity and review interfaces.

An agent workflow with explicit review

Orca can explore a project, maintain a task checklist, use multiple tools, and continue working through a multi-step request. Its two operating modes keep planning and mutation clearly separated:

• Plan mode is read-only for project and external state.
• Work mode can prepare changes and run supported workflows, but every model-requested project change waits for an explicit Apply or Reject decision.

Proposed file edits are hash-bound, validated, checked for unsaved or independently changed files, and presented as readable diffs. Apply performs the safety checks again before writing. A guarded revert remains available while its in-memory checkpoint is valid and the target has not changed independently.

Focused run and verification support

In Work mode, Orca can launch a saved current or main scene in one bounded process that it owns. It captures limited standard output and error evidence and can evaluate narrowly defined startup or expected-exit criteria. Orca does not claim that process output proves visual or gameplay correctness.

Open source and provider-independent

Orca is released under the MIT License. Its source code, provider integrations, tool contracts, safety boundaries, documentation, and automated tests are available for inspection and contribution.

You can choose from supported hosted providers, configure a compatible endpoint, or use local inference. Orca is not tied to one model vendor.

Highlights

• Streamed responses and tool calls with real cancellation.
• Request-scoped awareness of the active scene, selection, script, caret, and relevant project state.
• Visible grouped tool activity with status, timing, details, and editor navigation.
• Rich code presentation and unified or expanded side-by-side change review.
• Explicit Apply, Reject, Revert, and Open actions for supported proposals.
• Persistent, project-scoped conversation history and multi-step task lists.
• Conservative context-window and session-cost visibility when provider metadata is available.
• Bounded project instructions and explicitly loaded project skills for teams that keep agent guidance with their project.
• Repetitive and no-progress tool-loop protection below the hard execution limits.

What's New In 1.2.1

Orca 1.2.1 is a reliability-focused release for long, stateful Godot workflows:

• Strict provider completion validation, bounded total-generation deadlines, and provider/turn lifecycle ownership.
• Explicit Plan-to-Work requests without weakening normal mutation approval.
• Atomic conservative tool batches, semantic loop detection, and safe context-pressure finalization.
• Deterministic mutation recovery outcomes with retained proposal and post-write verification.
• Exact same-turn game-process cleanup, bounded final evidence, and a tested run/fix/rerun/verify workflow.
• Privacy-safe support diagnostics, resilient active composer behavior, and corrected Expanded Diff window ownership.

Complete release details are available in the changelog or download the latest build from GitHub Releases.

Local Models

Orca can connect to chat-capable models served by Ollama, LM Studio, and other local OpenAI-compatible runtimes. Local inference gives you greater control over model selection and where requests are processed while retaining Orca's in-editor workflow.

Local and editable profiles begin in Chat mode. Before Orca exposes project Agent tools, an isolated two-step function-call probe must pass for the exact provider, origin, endpoint, model, reasoning configuration, and probe version. You must then explicitly enable Agent tools for that configuration.

The compatibility probe receives no project files, editor context, project instructions, skills, task list, conversation history, or real Orca tool schema. Passing the probe confirms one narrow protocol exchange; it does not guarantee that a model will plan effectively or use tools reliably on a real project.

Orca does not install or download models, bundle an inference engine, or manage Ollama or LM Studio processes. The local server remains under your control. A local endpoint may still forward received data elsewhere, so use only software and endpoints you trust.

Supported Providers

• OpenAI
• Google Gemini
• xAI
• DeepSeek
• OpenRouter
• Ollama
• LM Studio
• Local OpenAI-compatible servers
• Custom OpenAI-compatible endpoints

Orca currently uses a streamed Chat Completions transport. Credentials, endpoint, model, and reasoning settings are retained separately for each provider. API keys are optional where the selected server does not require authentication.

Safety Model

Orca is designed to assist with development work without silently taking ownership of the project:

• Project access remains inside res://, rejects symbolic-link traversal, and excludes Orca's own add-on directory from agent tools.
• Existing files require the exact content hash returned by Orca's file reader before a change can be proposed.
• Known unsaved scripts and scenes are protected from reviewed writes.
• Supported GDScript changes are validated before review and immediately before writing.
• Apply rechecks file identity, hashes, editor state, proposal integrity, and validation.
• Searches, reads, tool rounds, responses, network inactivity, and Orca-owned game processes are bounded.
• Provider settings are snapshotted for each turn so a tool loop cannot silently switch endpoint or model.

These safeguards complement normal development practices. Orca does not replace source control, code review, backups, or project testing.

Installation

1. Download Orca from the Godot Asset Library or obtain the latest release archive from GitHub.
2. Ensure the add-on is installed at addons/orca/ inside your Godot project.
3. Open the project in Godot 4.7.2.
4. Enable Orca under Project > Project Settings > Plugins.
5. Open the Orca dock, select a provider, configure a model, and start a session.

Latest releases: https://github.com/cold-atom/Orca/releases

Data and Operational Notes

• Supported Godot version: Godot 4.7.2. Broader compatibility is not currently claimed.
• Prompts, relevant editor context, and project content returned by tools may be sent to the provider you select. Use only providers and endpoints you trust.
• Editable non-loopback endpoints require confirmation for their exact scheme, host, and effective port. Confirmation does not prove that a server is trustworthy.
• API credentials are saved through Godot Editor Settings. This is not an encrypted operating-system credential vault.
• Project conversation history is stored locally as plaintext JSON under Godot's user:// storage.
• Pending approvals and revert checkpoints do not survive an editor or plugin restart.
• Structured scene operations intentionally support a constrained set of nodes, values, dependencies, and signals.
• Runtime verification is limited to bounded process evidence and predeclared startup or exit criteria. It cannot establish visual or gameplay correctness.

Project Status

The current release is Orca 1.2.0 for Godot 4.7.2. Orca is a functional development-stage agent with permanent automated tests and continuous integration, but it is still evolving. Review proposed changes carefully and use version control on real projects.

Source and Documentation

Orca is open-source software released under the MIT License.

Source code, documentation, releases, and issue tracking: https://github.com/cold-atom/Orca

Contributions and constructive feedback are welcome.