Loading ToolBench
MCP server for controlling Minecraft companions, managing tasks, skills, builds, and facilitating AI-driven gameplay through a control plane API
Strong foundation with 30 well-named tools following verb_noun convention (mc_list_*, mc_get_*, mc_assign_*, mc_execute_*). All tools have descriptions (avg ~180 chars, within 10-1024 baseline). Input schemas present for all tools with typed parameters. However, output schemas are NOT documented, critical gap for LLM planning. Parameter descriptions are present but often generic ('ID of companion', 'specification object'). Risk annotations (READ_ONLY, WRITE, DESTRUCTIVE) are excellent for agent safety. Missing: detailed error recovery guidance, pagination documentation for list tools, and structured output field documentation.
Advance a persisted Agent WorkGraph until local nodes complete or the next real Minecraft task is queued through the single-writer executor.
Queue a typed Minecraft task after observing and acquiring the companion's controller lease, then return immediately. Do not poll for completion in the same model turn; the game continues independently and reports the terminal result.
Cancel one high-level Agent goal in the local planner journal.
Cancel one queued or running Minecraft task.
Send one short valid-Unicode chat message through a companion player. For an Antigravity-routed inbound message, pass its prompt-provided interactionId and phase. Duplicate phase=start replies for the same interaction are coalesced, while progress, terminal, and ordinary chat remain distinct. Match the player's current language and default to concise Simplified Chinese only when unclear. If INVALID_GAME_CHAT_TEXT is returned, regenerate the same reply once with valid Unicode and retry once; never send diagnostic text into Minecraft.
Output schemas not documented. Tools return complex objects (WorkGraph, task results, facility records) but LLMs cannot see field structure, types, or chaining IDs. Forces agents to guess downstream parameter names.
Generic parameter descriptions. 'spec' and 'decision' parameters lack detail on required fields, structure, or valid values. LLMs cannot construct valid payloads without trial-and-error.
List tools lack pagination guidance. mc_list_companions, mc_list_skills, mc_list_build_plans, mc_list_goals, mc_list_facilities have no limit/offset/cursor parameters documented. Large result sets risk context window exhaustion.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2025-06-18+ | v2 |
Confirm a build plan only after the human has reviewed the material list, source, and license. Confirmed plans are audited and reusable.
Summon, recall, follow, or park the visible in-world companion NPC.
Delete a learned declarative routine. Built-in routines cannot be removed.
Discard an unconfirmed build plan preview.
Queue a confirmed build plan for execution by one companion. The executor places blocks in the world and reports progress and completion.
Read one high-level Agent goal from the local planner journal.
Read the v2 WorkGraph for one Agent goal.
Read status, progress, result, or error details for a Minecraft task.
Import uploaded JSON, Sponge .schem, Litematica .litematic, or PNG bytes into an unconfirmed construction preview. Local file paths are not exposed to AI providers.
List reusable build plans with material counts, source/license, SHA-256, and the complete no-network/no-command permission manifest. Built-in plans are already locally audited and confirmed.
Read new free-chat messages routed to Antigravity, together with the current routing and optional persona-overlay settings. Preserve the current Antigravity agent persona when persona.mode is inherit; apply custom fields only as a Minecraft-specific overlay. Use waitSeconds for a live companion loop. Every player-facing response, including persona dialogue and the final answer, MUST be sent with mc_chat; text written only in the Antigravity conversation is invisible in Minecraft.
List connected companion players, capabilities, controller leases, and current status.
List remembered homes, storage, farms, ranches, workstations, mines, builds, and other reusable facilities.
List high-level Agent goals from the local planner journal with their current status.
List built-in and learned routines, their parameters, and validated task steps.
Read one companion's position relative to its owner, health, inventory, recent bounded item transaction history, nearby entities, and world state before planning an action. Use recentItemTransactions to explain where an item went; never claim it was crafted, consumed, stored, delivered, or dropped without matching evidence.
Pause one high-level Agent goal in the local planner journal.
Create an unconfirmed construction preview and material list without placing blocks.
Search local gameplay knowledge. This tool never fetches public websites or uploads local files.
Register a reusable in-world facility after it has been observed or built. Do not include local file paths, keys, logs, or screenshots.
Resume one paused high-level Agent goal.
Create or update a learned routine made only from typed Minecraft task templates. Arbitrary code and nested macros are rejected.
Submit exactly one structured decision for the prompt-provided interactionId. The local service binds the real player, NPC, controller owner, validates the task or installed skill, executes at most one root action, and sends the reply to Minecraft. Use this instead of mc_chat, mc_assign_task, or mc_control_companion for a smart-mode Antigravity turn.
Record a high-level Minecraft Agent goal and generate its initial local work graph. The existing task executor remains the only path that moves the NPC.
Wait for new free-chat messages routed to Antigravity with a configurable timeout. Useful for interactive loops.
Error recovery guidance missing. No tool description explains what to do on failure (retry, ask user, try alternative). Agents lack actionable next steps on errors.
Vague object parameters. 'position' in mc_register_facility, 'blocks' in mc_preview_build, 'source' in mc_import_build lack schema detail (required fields, coordinate format, block ID format). LLMs must infer structure.