MCP server exposing the live Unity Editor to an AI CLI (Claude Code, Antigravity, or Codex) over a localhost TCP bridge: reads console, compile errors, project info and the scene, and drives GameObjects, components, assets, play mode and the Test Runner.
Agen-Link exposes 13 tools with detailed descriptions (avg 250+ chars) and clear action verbs (agen_ping, agen_get_*, agen_refresh_*, agen_find_*, agen_graph_*). Most tools have well-documented purposes and prerequisites. Output schemas are entirely undocumented, LLMs cannot predict return structures. Error handling is absent: no guidance on retryability, user-fixable vs fatal errors, or recovery paths. Parameter descriptions exist for tools with inputs but lack format/range constraints (e.g., agen_read_console.max accepts 1-500 but no validation guidance). Tool composition is sound (each does one thing), but chaining is implicit, no documentation of which tools' outputs feed into others' inputs.
Search the project's asset database with a Unity filter string (e.g. 't:MonoScript', 't:Prefab', 't:Material wood', 'PlayerController t:MonoScript', 'l:Player'). Returns matching asset paths and GUIDs. Read-only — to move/copy/delete or create assets use agen_manage_asset.
Get the current C# compile errors and warnings from Unity's latest compilation. ALWAYS call this after creating or editing scripts (following agen_refresh_assets): if errorCount > 0, read the messages and fix the code before finishing.
Get the currently open Unity project's info: Unity version, absolute project path, product/company name, active build target/platform, render pipeline, active scene path, loaded scenes, and play/compile state. Check this instead of asking the user which scene is open or whether the game is running.
Get the active scene's GameObject hierarchy: instanceID, name, hierarchy path, active state, attached component type names, and children, down to a depth. The instanceIDs it returns can be passed straight to agen_get_gameobject / agen_modify_gameobject / agen_manage_component. ALWAYS call this (or agen_find_gameobjects) instead of asking the user what is in the scene.
Get the objects currently selected in the Unity Editor (names, types, and asset paths if any). Read-only — to change the selection use agen_set_selection.
6 tools have empty input schemas (agen_ping, agen_get_project_info, agen_get_compile_errors, agen_refresh_assets, agen_get_selection, agen_graph_build, agen_graph_status). LLMs cannot validate inputs or understand parameter constraints.
No output schemas documented for any tool. LLMs cannot predict return structures, field names, or types. This forces agents to guess at response shapes and breaks tool chaining (e.g., agen_graph_query returns 'node IDs + typed edges' but structure is opaque).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
(Re)build the cached dependency/knowledge graph of the open Unity project: scripts (inheritance, interfaces, serialized-field composition) and prefabs/scenes (attached MonoBehaviour components + asset references). Returns immediately; the build runs in the Editor — then poll agen_graph_status until building=false and hasCache=true. NOTE: this is a STRUCTURAL graph, not a method-call graph.
Assign human-meaningful names to one or more Neuron systems (clusters). Pass the system `id`s from agen_graph_systems and a concise name for each (e.g. 'Teleport Locomotion', 'Gaze Interactio
Query the cached Unity project dependency graph — the live WIRING you can't easily grep: which scripts are components on which prefabs/scenes, prefab/scene -> asset references, and script inheritance / interface / serialized-field links. Two modes: (1) FOCUSED — pass `entity` (script full name, asset path, GUID, or display name) to get that node + its neighbors within `depth` hops; (2) WHOLE-PROJECT FILTERED — omit `entity` and pass `kinds`. Returns node IDs + typed edges only (no file contents), capped by `limit`. If the cache is empty, call agen_graph_build first and poll agen_graph_status.
Check the Unity project knowledge-graph cache: is it building?, does a cache exist?, and node/edge counts. Poll this after agen_graph_build until building=false and hasCache=true before agen_graph_query.
List the Neuron graph's auto-detected 'systems' (clusters of inter-linked scripts/assets, grouped per scene plus a Shared·Core and a Project bucket). Each entry has: id, current name, owner (scene/shared/project), the main (hub) script, member names, and `needsNaming` (true when the cluster has no cached human name yet). Use this to find clusters that need naming, then call agen_graph_name_systems. Naming is cached by cluster membership, so already-named clusters stay named across rebuilds with no LLM calls.
Health probe. Answers from the bridge's socket thread WITHOUT touching Unity's main thread, so it still replies while the Editor is busy. Call this FIRST whenever another agen_* tool times out — it tells you which half is actually broken: • no response at all → the bridge really is down (Unity closed, or the package failed to load). • ok with mainThreadResponsive:true → bridge and Editor are both fine; that one command is slow. • ok with mainThreadResponsive:false → the bridge is fine and Unity's main thread is BLOCKED (asset import, shader compile, modal dialog or progress bar) or the Editor is parked. Do NOT report this as a bridge failure, do NOT tell the user to click the Editor if it is blocked, and do NOT fall back to writing editor scripts — read 'verdict', tell the user what Unity is busy with, and retry.
Read recent Unity Editor console messages (errors, warnings, logs). Use this to see runtime problems, e.g. after the user runs the game, or to review a project for issues. Filter by type and limit the count. ALWAYS check the console here (and agen_get_compile_errors) yourself before asking the user about errors.
Ask Unity to import changed files and recompile scripts (AssetDatabase.Refresh + RequestScriptCompilation). Call this after writing/editing C# files. It returns {scheduled:true} immediately, BEFORE the refresh runs — deliberately, because a script change triggers a domain reload that would otherwise kill this reply in flight. So the isCompiling it returns is the OLD value; do not read anything into it. Then poll agen_get_compile_errors until isCompiling is false and read errorCount. The bridge may drop for a few seconds during the reload and reconnects itself — a failure there means retry, not that the refresh failed.
No error handling guidance. Tools lack recovery hints (e.g., 'If agen_refresh_assets times out, call agen_ping first'). No categorization of errors as retryable vs fatal. agen_ping's description mentions 'verdict' field but no schema documents it.
agen_graph_name_systems description is truncated ('Assign human-meaningful names to one or more Neuron systems (clusters). Pass the system `id`s from agen_graph_systems and a concise name for each (e.g. 'Teleport Locomotion', 'Gaze Interactio'). Incomplete docstring prevents LLM from understanding full intent.
Tool composition lacks explicit chaining documentation. agen_graph_build → agen_graph_status → agen_graph_query is a required sequence, but no tool description states 'Call agen_graph_build first, then poll agen_graph_status until building=false before calling agen_graph_query.' Agents must infer dependencies.