m4bwav/unity-agent
Evergreen skills for driving Unity with AI coding agents: the official Unity CLI and Pipeline package, headless batchmode and compile checks, MCP bridge selection, agent-safe Unity workflow rules, and UPM package management. Each skill re-researches its subject on a schedule and carries the evals that prove it triggers and acts.
Drive Unity from the terminal with the official Unity CLI (the unity binary, beta since July 2026) and its com.unity.pipeline package: list or install editors and modules, open or close a project, see which Editors are running, run EditMode/PlayMode tests with a results report, build players, execute registered Editor commands or C# eval against a running Editor, and start or configure Unity's built-in MCP server. Use whenever the user asks to control, script, automate, or talk to Unity, Unity Hub, or the Unity Editor from the command line, mentions the Unity CLI, unity status, unity test, unity build, unity command, the Pipeline package, or asks 'can Claude drive the editor', 'run the Unity tests', 'which editors are installed', 'is the editor open', 'install Unity 6.x', or 'refresh unity-agent-cli' / 'is unity-agent-cli stale'. Batchmode work with no CLI is the sibling unity-agent-headless skill; choosing an MCP server is unity-agent-mcp.
Run Unity without opening the Editor window and prove the result from logs: batchmode -executeMethod passes, Unity Test Framework runs from the command line (-runTests, NUnit XML), an offline C# compile check that works even while the Editor is open (dotnet build on Unity's generated csproj using the .NET SDK bundled inside every Unity install), Editor and build log locations, exit-code traps, the project lockfile, and the deprecated Unity Hub headless CLI as a fallback. Use whenever the user asks to build, test, import, or run something in Unity headlessly, in batch mode, from a script or CI, wants to know if the code compiles without opening Unity, asks where the Editor log is, why a batchmode run returned 0 with errors, why 'dotnet test' says no SDK, or says 'refresh unity-agent-headless' / 'is unity-agent-headless stale'. Driving a running Editor through the official CLI is unity-agent-cli; choosing an MCP server is unity-agent-mcp.
Choose, install, configure, and troubleshoot an MCP server that lets an AI agent (Claude Code, Claude Desktop, Copilot, Cursor, Codex) see and edit a running Unity Editor: Unity's official unity mcp mode (from the Unity CLI and com.unity.pipeline), CoplayDev MCP for Unity, IvanMurzak Unity-MCP, CoderGamester mcp-unity, docs servers, and when to use MCP versus the CLI versus code-only work. Use whenever the user asks 'is there an MCP server for Unity', 'connect Claude to Unity', 'let the agent see the scene', 'add unity-mcp', 'which Unity MCP should I use', 'configure MCP for the editor', reports the Unity MCP bridge dropping after a domain reload or play mode, or says 'refresh unity-agent-mcp' / 'is unity-agent-mcp stale'. Running CLI commands against the Editor once a route is chosen is unity-agent-cli; batchmode and compile checks are unity-agent-headless.
Add, remove, pin, upgrade, or look up Unity Package Manager (UPM) packages from outside the Editor UI: editing Packages/manifest.json safely, the Unity Package Manager Client API run headless (the async poll pattern without -quit), the Pipeline package_add/package_remove commands on a running Editor, OpenUPM and git-URL packages, packages-lock.json, and which AI-related Unity packages are safe on which Unity version. Use whenever the user asks to install or remove a Unity package, add com.unity.something, bump a package version, fix a package resolution error, add an OpenUPM or GitHub package, or says 'refresh unity-agent-packages' / 'is unity-agent-packages stale'. General CLI use is unity-agent-cli; batchmode mechanics are unity-agent-headless.
How an AI coding agent works inside a Unity project so its changes are verifiable: the verification ladder (engine-free tests, offline compile check, Unity Test Framework, live Editor checks), the rules that keep Unity repos safe for agents (never hand-write scene/prefab/asset YAML, keep .meta files with their assets, Humble Object logic, importers over hand-edited assets, domain reload), repo setup for agents (AGENTS.md snippet, skill packs, .gitignore, csproj generation) and Unity's stance on AI-made content. Use whenever the user asks how Claude, Copilot, Cursor or Codex should work with Unity, wants a Unity repo set up for AI agents, asks for Unity coding conventions for agents or the latest thinking on AI with Unity, asks why an agent edit broke a scene or .meta, or says 'refresh unity-agent-workflow' / 'is unity-agent-workflow stale'. The CLI is unity-agent-cli; batchmode and compile checks are unity-agent-headless; MCP servers are unity-agent-mcp; packages are unity-agent-packages.