UnitySkills lets an AI coding agent operate the Unity Editor. A Unity Editor plugin runs an HTTP server that exposes 805 skills as REST endpoints across 54 categories, and an agent skill package (a skill definition, a Python client, and module docs) teaches the agent when and how to call them. It is a refactor and extension of the Unity MCP concept, and it targets Unity 2022.3 and Unity 6.
The skills cover most Editor work: game objects, components, prefabs, scenes, materials, shaders, Shader Graph, UI and UI Toolkit, animation, Timeline, physics, lighting, terrain, navigation meshes, assets and import settings, builds, tests, profiling, and project validation. Package-specific modules drive Cinemachine, Netcode for GameObjects, XR, ProBuilder, URP, Addressables, YooAsset, HybridCLR, DOTween, and QFramework. 28 advisory design modules on architecture, performance, and maintainability load on demand, and each skill module carries do-not lists and routing rules for the agent.
A governance layer covers the call lifecycle. Dry-run and plan modes validate parameters and estimate impact without writing anything. A server-side permission system with Approval, Auto, and Bypass modes (aligned with Claude Code permission modes) gates high-risk operations from each skill's risk metadata, with a per-skill allowlist. Every call, grant, revoke, and block lands in a JSONL audit trail, and persistent snapshots survive domain reloads so one task can be rolled back without rewinding the project.
One-click installers in the Unity panel set up the skill for Claude Code, Antigravity, Codex, Cursor, OpenCode, and Kimi Code, and any tool that reads standard agent skill directories can load it.
Features
- 805 REST skills: Editor operations across 54 categories, most with batch variants
- Batch transactions: multi-step batches with cross-step references, rollback on failure, and a net-change diff
- Permission modes: Approval, Auto, and Bypass, with high-risk gating from skill risk metadata and a per-skill allowlist
- Dry run and plan: parameter validation and impact estimates before anything is written
- Audit trail: structured JSONL record of every call, grant, revoke, and block, browsable in the Unity panel
- Per-task rollback: five snapshot types that persist across domain reloads and sessions
- Multi-instance control: a global registry discovers and controls several Unity projects at once
- Long connections: configurable request timeouts and automatic recovery after a domain reload
- UnityCLI binding: cold-starts the bound project without Unity Hub
- Advisory modules: 28 design guides on architecture, performance, and package-specific rules
- Agent installers: one-click setup for Claude Code, Antigravity, Codex, Cursor, OpenCode, and Kimi Code
