MCP server that bridges a local Minecraft client to ChatGPT Desktop, providing tools to query game state, read files, manage skills, and communicate with players
MineGPT Bridge defines 23 tools with consistent naming (verb_noun pattern: get_, list_, minegpt_*) and reasonable descriptions (avg ~120 chars). All tools have input schemas with typed parameters and descriptions. However, output schemas are not documented in the source code, LLMs cannot predict return structures. Error handling is absent: no recovery guidance, no categorization of retryable vs fatal errors. Parameter descriptions lack constraints (e.g., minegpt_read_game_file offset/length have no bounds; minegpt_get_nearby_entities radius defaults to 32 but max is 64, not stated in description). No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. Composition is sound: tools chain well (list_skills → get_skill, list_mods → inspect_mod_jar). Security: no secrets exposed as params, but no audit trail or permission gates documented.
Read the player's current biome, time, weather, difficulty, and local light levels.
Read one loaded block's ID, state, light, and block-entity type. This never loads a chunk.
Read a 16 by 16 surface summary of one chunk already loaded by the Minecraft client. Omit both coordinates for the player's current chunk. This never loads a chunk.
Read block-ID counts for one loaded 16 by 16 by 16 chunk section. This is a summary, not a full block dump.
Read and parse the active instance's options.txt client settings.
Read non-empty player inventory, armor, and offhand slots without item NBT.
Output schemas not documented. LLMs cannot predict return structures for any of the 23 tools. This forces agents to guess what fields are available and risks failed downstream tool calls.
No error handling guidance. Tools lack recovery hints (e.g., 'File not found. Check minegpt_list_game_files first.'). No error categorization (retryable vs fatal). Agents cannot self-correct on failures.
Parameter constraints missing. minegpt_read_game_file (offset, length) and minegpt_get_nearby_entities (radius) lack bounds. minegpt_get_recent_log (lines) has no max specified. LLMs can pass absurd values causing timeouts or errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Read up to 64 client-visible entities near the player. Radius defaults to 32 and is limited to 64 blocks.
Read the player's current position, health, hunger, experience, game mode, and dimension.
Read the tail of logs/latest.log from the active instance, up to 1000 lines.
Read one Markdown skill from the local Minecraft minegpt/skills directory. Call minegpt_list_skills first and pass its relative path.
Read the block or entity under the player's crosshair, if any.
Download one explicitly requested Markdown skill from a public GitHub repository into the local Minecraft minegpt/skills directory. Existing files are not overwritten unless overwrite is true.
Inspect one direct mods/*.jar file without executing it. Returns bounded Mod metadata plus matching recipe/resource text or printable class strings.
Recursively list files and directories in the active Minecraft game directory without leaving it. Results are capped at 500 entries.
List .jar files directly in the active instance's mods directory.
List direct child world directories in the active instance's saves directory.
List user-editable MineGPT Markdown skills, up to eight nested directories deep, from the local Minecraft minegpt/skills directory.
Wait up to 45 seconds for the next Minecraft player message.
Return the one-time pairing token for the Minecraft client.
Read a bounded byte range from a regular file in the active Minecraft game directory as UTF-8 text or Base64.
Send a reply to an outstanding Minecraft player message.
Search bounded text files in likely recipe, KubeJS, config, datapack, and FTB Quests locations of the active instance. This is literal, read-only local evidence only.
Return MineGPT bridge connection and queue status.
No tool annotations. Despite clear risk levels (READ_ONLY vs WRITE), tools lack readOnlyHint/destructiveHint/idempotentHint. Agents cannot reason about side effects or retry safety.
Ambiguous tool name: minegpt_get_biome_and_environment combines two concerns (biome + time/weather/difficulty). Should split into minegpt_get_biome and minegpt_get_environment for single responsibility.