UnitySkills

🎮 UnitySkills

  REST API-based AI-driven Unity Editor Automation Engine
  Let AI control Unity scenes directly through Skills

  🎉 We are now indexed by DeepWiki!
  Got questions? Check out the AI-generated docs → 

> The current official maintenance baseline is Unity 2022.3+. Some Unity 2021 compatibility logic may still remain in the codebase, but future feature work, regression testing, and adaptation will focus on 2022.3+ / Unity 6.

💖 Sponsors

    Sponsor
    Contribution
    Provides substantial API credits for the upcoming v3.0, with continued API support to follow.

  Become a sponsor →

📈 Project Contribution Rankings

🤝 Acknowledgments
This project is a deep refactoring and feature extension based on the excellent concept of unity-mcp.

🚀 Core Features

- 🛠️ 839 REST Skills Toolkit: 56 source files across 54 categories, plus 29 advisory design modules, with Batch operations.
- ⚡ UnityCLI Support: Bind UnityCLI to cold-start the bound project — no need to launch Unity Hub.
- 🔐 Three-Tier Permission Modes: Approval / Auto / Bypass with dual approval channels, aligned with Claude Code permission modes.
- 🤖 6 Major IDEs Native Support: Claude Code / Antigravity / Codex / Cursor / OpenCode / Kimi Code — one-click install and use.
- 🛡️ Transactional Atomicity: Failed operations auto-rollback, leaving scenes clean.
- 🌍 Multi-Instance Control: Automatic port discovery and global registry for controlling multiple Unity projects at once.
- 🔗 Stable Long Connections: Configurable request timeout (default 15 minutes), automatic recovery after Domain Reload.
- 🛡️ Anti-Hallucination Guardrails: Each Skill module includes DO NOT lists and routing rules.

🛡️ Built for Trust: The Governance Layer

An AI driving the Editor writes to real scenes, prefabs and .meta files. The interesting question isn't whether it *can* — it's what happens when it gets something wrong. UnitySkills answers that at four points in the call lifecycle:

- Before execution: ?mode=dryRun / ?mode=plan rehearsal — returns parameter validation and an impact estimate, writes nothing.
- At execution: the server derives high-risk gating (NeverInSemi) from each skill's risk metadata — gating never depends on the AI behaving well; Allowlist grants a persistent per-skill pass.
- After execution: every call, grant, revoke and blocked hit lands in a JSONL audit trail, browsable in the panel; deletions are audited too.
- After a mistake: five types of persistent snapshots survive Domain Reloads; workflow_undo_task rolls back one task, not the whole project; POST /skills/batch runs as a transaction (cross-step $ref, rollback on failure, ?diff=1 net change).

Full mechanics → Operating Modes & Governance

How this compares

| Dimension | UnitySkills | Typical MCP bridge | Unity's official AI Assistant |
| Permission granularity | Per-operation: three modes (Approval / Auto / Bypass) + risk metadata on every skill + per-skill Allowlist | No permission model — once connected, the whole tool surface is callable | Per-client trust (Pending Connections → Allow / Revoke); after approval, calls are not distinguished by operation |
| Audit | Structured JSONL for every call / grant /

……（后略，见原文）