Unity Explorer MCP Scene Iteration

Drive a running Decentraland Explorer build through its MCP automation server to build and test SDK7 scenes autonomously: edit the scene, watch it hot-reload, move the camera and player, take screenshots, and verify against what the code should produce.

The connected mcp__explorer__* tools are self-describing — each carries its name, arguments, and output shape. Treat that as the authoritative tool catalog; the names used below (get_scene_state, get_scene_logs, screenshot, teleport, move_to, walk, look_at, set_camera_pose, set_camera_mode, list_scene_entities, get_entity_details, get_player_state, click_entity, send_chat, trigger_emote, reload_scene, get_scene_content_stats, get_scene_content_breakdown, get_performance_stats) are the common ones.

Coming in through the Creator Hub MCP? When the Creator Hub's editor MCP is connected (skill creator-hub-mcp; tools mcp__creator-hub__*), the Explorer is reached through it: launch_preview starts the scene preview with this MCP server on, and the runtime tools below are re-published live as explorer_<name> (explorer_screenshot, explorer_walk, …; explorer_call(tool, arguments) is the fallback while they bind). In that case skip Setup below entirely — no claude mcp add, no bind gate, no port 8123 probe — and go straight to the iteration loop with those tool names. The reference files apply unchanged.

Deeper reference, loaded only when the task reaches it:

• reference/camera-and-movement.md — before framing screenshots, free-camera sweeps, or navigating precise lines
• reference/assets.md — before placing, downloading, converting, or exporting any 3D model
• reference/visuals.md — before tuning emissives/bloom, UI overlays, skybox time, or judging thin geometry
• reference/performance-debugging.md — before answering whether a scene is within its content budget, why it runs slow, or what to optimize (get_scene_content_stats, get_scene_content_breakdown, get_performance_stats)

Gates

Certain points in this skill are gates: you ask, call no tool after asking, and let the user answer. A gate opens only on their reply — never on your own judgment, never on their silence, never because a workaround is available to you. Running as a subagent, you cannot open a gate at all — there is no user to ask: stop and report the pending decision to your caller with your recommendation, rather than passing the gate on your own authority. Doing so violates the gate even when the workaround happens to work. The gates, in order: the skills-install gate and restart gate (below), the intent gate (pre-flight), the launch/kill gate (Setup step 1), and the bind gate (Setup step 2).