?
Host a serverModrinth App
Sign in
ModsResource PacksData PacksShadersModpacksPluginsServers
Settings

VitaminMCP

An MCP server that enables AI agents to perform real end-to-end testing of Minecraft servers using bots and server introspection.

89downloads
0followers
89downloads
0followers

Links

Report issues View source Visit wiki

Tags

Creators

Details

Licensed MIT
Published 4 days ago
Updated 5 days ago
All versions

3.0.3

Release
3.0.32 months ago 5
Download

Compatibility

Minecraft: Java Edition
1.21.x
Platforms
Paper
Purpur

Changes

Token-usage optimization for MCP clients, multi-client install docs, and a test fix.

Changed

  • Tool results are serialized as compact JSON instead of pretty-printed, on both the agent endpoint and the local MCP server.
  • Serialized records omit null fields (an event with no player, a log line with no throwable). Fields where null is meaningful — bot_inspect's menu/items — are unchanged.
  • Tool descriptions are ~29% shorter. Operational warnings are kept; the long walkthroughs live in the docs and the bundled skill.
  • wait_for's condition, state_query/inventory_contains' which, bot_view's what/mode and log levels are now schema enums instead of prose lists. The condition enum also documents player_state, which existed but was missing from the description.
  • Default response budget per query page lowered from 50 KB to 25 KB (still 200 records); every query already pages with nextCursor.

Added

  • config.yml: max-response-items and max-response-bytes to tune the response budget.
  • README: per-client MCP registration guide (Claude Code, Cursor, Codex CLI, Gemini CLI, Windsurf, Claude Desktop, VS Code), the setup prompt's actual command name for plugin installs (/mcp__plugin_vitaminmcp_vitaminmcp__setup), and connect-prompt variants that keep the token in a file or in VITAMINMCP_TOKEN instead of the conversation.
  • Version support table marks 1.18–1.20.6 as planned.

Removed

  • Pretty-printing of tool results.
  • Null fields from serialized records.
  • ~4,400 characters of tool-description prose.

Token savings vs 3.0.2

Where 3.0.2 3.0.3 Saving
Tool descriptions, session tools (every conversation) 7,352 chars 5,220 chars −29%
Tool descriptions, agent tools (per session_start) 8,317 chars 6,019 chars −28%
Large query responses (logs_query, events_query) pretty JSON + null fields compact, nulls omitted −15–25%
Response budget default (per page) 50 KB 25 KB configurable

A typical plugin-verification session carries roughly 20–25% less MCP-related context (~6k tokens). Cumulative billed savings grow with session length, since every response is re-read on each later turn.

Fixed

  • BotRunnerCommandTest asserted Windows' node.exe and failed on macOS/Linux; it now asserts against the host's own runtime.
  • npm README claimed support up to 1.21.8; the supported ceiling is 1.21.11.