Skip to main content
Glama
BlockbenchMCP
by sosadly

Blockbench MCP

An MCP server for Blockbench: an AI models, textures, rigs and animates Minecraft models inside the editor.

MCP Blockbench Node License: MIT Stars

Quick start · Gallery · Features · Tools · FAQ


Blockbench MCP is an open-source Model Context Protocol server that connects AI assistants such as Claude Code, Claude Desktop and Cursor to the live Blockbench desktop editor. The AI can create Minecraft Java, Bedrock and GeckoLib models, paint textures pixel by pixel, build rigs, write keyframe animations, install Blockbench plugins and take screenshots to check its own work. You don't have to touch the editor.

It has 86 MCP tools over one local connection. They cover blockout, detailing, texturing, rigging, animation and export, with quality gates between the stages so a model ends up with real detail instead of 15 flat boxes.

All three models were built, textured and posed by an AI through Blockbench MCP.

Related MCP server: Blockbench MCP

Quick start

git clone https://github.com/sosadly/blockbench-mcp.git
cd blockbench-mcp
npm install && npm run build
  1. In Blockbench desktop, open File ▸ Plugins ▸ Load Plugin from File and pick plugin/blockbench_mcp.js. Allow network access when asked.

  2. Add the server to your AI client. Here is Claude Code as an example:

    claude mcp add blockbench -- node /absolute/path/to/blockbench-mcp/dist/index.js
  3. Ask for something, for example: "Make a textured GeckoLib bear that can walk, run, sleep and attack."

The full setup is under Installation.

How it works

There are two pieces:

Piece

Runs where

Responsibility

Bridge plugin (plugin/blockbench_mcp.js)

Inside Blockbench (desktop)

Hosts a local HTTP endpoint on 127.0.0.1:8787 (built on Node's net module) and runs each command against the Blockbench API on the renderer thread.

MCP server (src/ → dist/)

A Node process your AI client launches

Exposes Blockbench as MCP tools (stdio transport) and forwards every call to the bridge.

┌────────────┐   stdio (MCP)   ┌──────────────────┐   HTTP 127.0.0.1:8787   ┌──────────────────────┐
│  AI client │ ──────────────▶ │  blockbench-mcp  │ ──────────────────────▶ │  Blockbench + plugin │
│ (Claude…)  │ ◀────────────── │   (Node server)  │ ◀────────────────────── │   (live editor)      │
└────────────┘                 └──────────────────┘                         └──────────────────────┘

Features

  • 🔎 References from the web. For a known character, creature, item, vehicle or product, find_references searches the web (safe search always on) and returns one numbered contact sheet plus a short description; load_reference {result:N} / {url} loads the chosen ones into Blockbench, and zoom_reference enlarges details like a face or a buckle.

  • 🎨 Palettes. make_palette reads a reference or concept image into named colour families, each with a 5-step hue-shifted ramp (or builds a harmony from a text brief), stores it on the project and lets texturing use color:"$hair" / "$hair.shadow".

  • 🧱 Modeling. Create bones (groups) and cubes one at a time or in bulk (add_groups / add_cubes build a whole posed skeleton in one call), with full rotation and inflate. Place cubes by center+size, attach them to a face of another cube (they land on the surface and follow its rotation and bone, so nothing floats), or build both sides at once with mirror. Edit, move, reparent and delete elements, and read the outliner tree.

  • ⚙️ Round shapes, mechanisms, cloth and hair. add_cylinder (solid or tube, an exact regular polygon), add_gear (teeth, spokes, hub, axle; mesh_with places the next gear so the teeth interleave), add_pipe, add_flare (pleated skirts, capes, lampshades, roofs), add_hair (bangs, side locks, back strands, tails on bone chains with ribbons), and animate_mechanism (spinning gear trains with the right ratios and directions, pistons, levers, pulsing lights, seamless loops).

  • ✨ No z-fighting. Coplanar faces are detected in world space (rotations, bones and inflate included) and fixed automatically for everything a tool creates; fix_z_fighting cleans up after manual edits.

  • 🏗️ Procedural detail. An LLM cannot compute 200 sets of [from,to] in its head, so these tools do the math. voxelize_matrix extrudes a character matrix into cubes (draw a blade, an emblem or a horn profile as pixel art). add_hollow_volume builds a shell with a real cavity (hoods, helmets, pauldrons, cages). generate_array repeats an element along a line, ring or grid with jitter, taper, rotation and an anti-z-fighting depth stagger (torn hems, scales, plates, teeth, rivets). extrude_chain builds a tapering, curving chain with one bone per segment (horns, tails, tentacles, braids). add_wing builds a jointed wing with a membrane.

  • 📏 Density gate. audit_complexity grades the model against a cube budget (prop 30-60, mob 100-180, hero 180-300+) and reports monolithic boxes, overlap, micro-detail density, bare flat faces and bone depth. The verdict (too_primitive / acceptable / high_detail) stops a blockout from reaching texturing.

  • 🖌️ Texturing engine. detail_cubes shades every pixel from its 3D point: hue-shifted light (cool shadows, warm lights), ambient occlusion, edge bevels and per-material detail (skin, cloth folds, hair strands with a shine band, metal sheen and scratches, wood grain, stone, leather stitching, scales, glowing emissive), plus world-space patterns that run across neighbouring cubes (plaid, stripes, panels with rivets, lace and trim borders, hazard, bricks…). paint_face draws eyes, brows, mouth and blush as pixel art in the face's own grid. pack_uv gives thin cubes per-face UV so every face gets real pixels. Low-level paint ops (pixels, rects, ellipses, polygons, dither, noise, gradients) and face-relative paint_faces are still there.

  • 🦴 Rigging. create_rig builds a segmented humanoid or quadruped skeleton with three bones per limb (so elbows and knees bend), joints on the real pivots and correct *_left / *_right naming. check_rig won't call a two-bone-limb rig animation-ready.

  • 🎬 Animation. generate_animation writes a complete base cycle (idle / walk / run / attack / cast / jump / hurt / death / fly) with correct gait phasing, bent joints, counter-rotation, follow-through and seamless loops. add_keyframes refines it in bulk with interpolation control.

  • 📐 Measured animation. analyze_animation evaluates the rig and reports how far each hand, foot and head travels in the model's own axes, whether the loop closes and whether lower limb segments move at all. This is how it catches an attack that swings into the model's back instead of at the target.

  • 🧭 Left and right stay correct. The model faces -Z, so its own right is +X. get_orientation, which_side and check_sides answer from coordinates. add_cube / add_group take side:"left"|"right" and refuse a call whose coordinate contradicts it, and every screenshot is labelled with which image edge is the model's right.

  • 🙋 Human review gate. request_review renders labelled views, shows them in the MCP Copilot panel and blocks until you press a button, then returns your verdict and comment inside the same tool call. ask_user does the same for a question.

  • 🎯 Reference matching. Drop a reference image into the MCP Copilot panel. compare_reference scores the silhouette match (IoU match_percent plus a [reference | model | overlay] composite and "add/trim mass here" advice), and measure_model gives numeric proportions.

  • 📸 Vision and close-up inspection. screenshot, screenshot_views and get_texture return images inline. check_model audits for z-fighting, floating parts, hidden (buried) cubes, stretched UVs and side mistakes; inspect_model (and the close-ups screenshot_views appends) frames each problem spot, x-rays away what stands in front, outlines the cubes involved and captions the fix — then checks the face, hands, feet, joints and gear contacts, which four full-body views never show.

  • 🧠 Guidance. get_guide returns playbooks for modeling, characters, mechanisms, detailing, inspection, orientation, rigging, texturing, VFX, animation, review and reference matching, and the server sends its workflow (references first, palette, build, inspect) to the client on connect.

  • 🪟 In-app panel. The MCP Copilot side panel shows server status, the reference drop zone, a live log of the AI's commands, model stats, the last match score and any pending review.

  • 🧩 Plugins. Search, install (by store id, URL or file) and uninstall Blockbench plugins, so the AI can set up formats like GeckoLib itself.

  • 📦 Export. export_project uses the format's own codec. export_model writes through any named codec; the default is glTF, a self-contained .gltf that imports into Godot, Unity or Blender.

  • 🔧 Escape hatch. execute_script runs arbitrary Blockbench JS for anything the dedicated tools don't cover.

Requirements

  • Blockbench desktop app 4.8+. The web app cannot host the bridge; see Limitations.

  • Node.js 18+ (for global fetch).

  • An MCP-capable client: Claude Code, Claude Desktop, Cursor, or any other client that speaks MCP over stdio.

Installation

1. Install the bridge plugin in Blockbench

  1. Open the desktop Blockbench app.

  2. File ▸ Plugins ▸ Load Plugin from File and select plugin/blockbench_mcp.js. You can also copy it into your Blockbench plugins folder.

  3. On first start the plugin asks for network permission because it needs the net module to host the local server. Choose "Always allow for this plugin."

  4. A toast confirms "MCP server started on port 8787."

    • Toggle the server any time: Tools ▸ Start / Stop MCP Server

    • Change the port: Settings ▸ General ▸ MCP Server Port (default 8787)

Verify it's up:

curl http://127.0.0.1:8787/ping
# {"ok":true,"protocol":1,"blockbench_version":"5.x.x","is_app":true,"has_project":false}

2. Build the MCP server

git clone https://github.com/sosadly/blockbench-mcp.git
cd blockbench-mcp
npm install
npm run build

3. Connect your AI client

Point your client at dist/index.js over stdio, using an absolute path.

Claude Code: add a .mcp.json to your project, or use claude mcp add:

{
  "mcpServers": {
    "blockbench": {
      "command": "node",
      "args": ["/absolute/path/to/blockbench-mcp/dist/index.js"],
      "env": { "BLOCKBENCH_MCP_PORT": "8787" }
    }
  }
}

Claude Desktop: claude_desktop_config.json. Cursor: .cursor/mcp.json. Both use the same shape:

{
  "mcpServers": {
    "blockbench": {
      "command": "node",
      "args": ["C:\\path\\to\\blockbench-mcp\\dist\\index.js"]
    }
  }
}

Env var

Default

Purpose

BLOCKBENCH_MCP_PORT

8787

Must match the plugin's port setting.

BLOCKBENCH_MCP_HOST

127.0.0.1

Bridge host.

Tool reference

Group

Tools

Status & guidance

get_status, get_guide, list_formats

Orientation (left/right)

get_orientation, which_side, check_sides

Human review

request_review, ask_user

References & colour

find_references, get_reference, load_reference, zoom_reference, make_palette, get_palette, compare_reference, measure_model, list_references, clear_references

Project

new_project, set_project_meta, save_project, export_project, export_model, load_project, close_project

Geometry

add_group, add_cube, add_groups, add_cubes, add_plane, add_mesh, add_meshes, mirror_element, edit_element, delete_element, list_outliner, get_element, pack_uv

Procedural detail

voxelize_matrix, add_hollow_volume, generate_array, extrude_chain, add_wing, add_flare, add_hair, add_cylinder, add_gear, add_pipe

Quality gates

audit_complexity, check_model, fix_z_fighting, inspect_model, check_sides, check_rig

Rigging

create_rig, check_rig, get_rig

UV & textures

create_texture, create_textures, create_vfx_texture, paint_texture, detail_cubes, paint_face, paint_faces, apply_texture, set_cube_uv, set_texture_render_mode, import_texture, resize_texture, list_textures, get_texture

Animation

animate_mechanism, generate_animation, analyze_animation, preview_animation, create_animation, add_keyframe, add_keyframes, list_animations, remove_animation

View

set_camera_angle, screenshot, screenshot_views

Plugins

list_plugins, install_plugin, uninstall_plugin

Escape hatch

execute_script

Conventions. Coordinates are Blockbench units, rotations are degrees, and texture pixel ops use a top-left origin with y pointing down. add_cube returns each face's resolved UV rect; detail_cubes then base-coats every face and paint_faces paints features (eyes, nose, claws) in coordinates relative to a face, so you never compute absolute UVs by hand.

Orientation. A model faces -Z, so its own right is +X and its left is -X. This is verified against Blockbench's bundled vanilla data: Bedrock's rightArm, stored at x = -5, is loaded by Blockbench at x = +5 (the Bedrock/Java codecs mirror X on import and export). Camera views are named from the model's point of view (front shows its face), and every capture is captioned with which image edge is the model's right, because a front view is mirrored just like facing a person.

Rotation signs. +X rotation lifts a bone's front, so a down-pointing bone (arm, leg) swings its tip forward while an up-pointing bone (torso, neck) tips backward. Elbows bend +X, knees bend -X, and +Y turns the model toward its own left. generate_animation applies these for you and analyze_animation verifies the result by measurement.

Examples

The sequence an AI follows for "make a textured GeckoLib bear that can walk, run, sleep and attack":

// 0. Read the playbook so the model comes out detailed and rotated, not boxy
get_guide {}

// 1. Install GeckoLib (adds the `geckolib_model` format)
install_plugin { "id": "geckolib" }

// 2. New project, straight from the start screen
new_project { "format": "geckolib_model", "name": "bear",
              "texture_width": 64, "texture_height": 64 }

// 3. Lay out the whole posed skeleton in one call; note the rotated bones
add_groups { "groups": [
  { "name": "body",   "origin": [0, 12, 0] },
  { "name": "head",   "origin": [0, 14, -7], "parent": "body", "rotation": [10, 0, 0] }, // nose down
  { "name": "leg_fl", "origin": [3, 8, -5],  "parent": "body", "rotation": [-6, 0, 0] },
  { "name": "leg_fr", "origin": [-3, 8, -5], "parent": "body", "rotation": [-6, 0, 0] }
  // ...hind legs, ears, muzzle, tail...
] }

// 4. Build many cubes at once (segment limbs, taper with inflate, mirror left/right)
add_cubes { "cubes": [
  { "name": "torso",   "from": [-5, 8, -7], "to": [5, 16, 7], "parent": "body" },
  { "name": "head",    "from": [-3.5, 10, -13], "to": [3.5, 16, -7], "parent": "head" },
  { "name": "muzzle",  "from": [-2, 10, -16], "to": [2, 13, -13], "parent": "head", "inflate": -0.5 }
  // ...
] }

// 5. Texture: shaded base coat on EVERY face (no gaps), then features face-relative
create_texture { "name": "bear", "width": 64, "height": 64, "fill": "#6e4a2b" }
detail_cubes   { "base": "#6e4a2b", "noise": 0.12, "bottom_dark": 0.3 }
paint_faces    { "faces": [
  { "cube": "head",   "face": "north", "ops": [
    { "type": "rect",    "x": 2, "y": 2, "width": 1, "height": 1, "color": "#0f0a05" }, // eye
    { "type": "ellipse", "x": 3, "y": 4, "width": 2, "height": 2, "color": "#140e08" }  // nose
  ] }
] }

// 6. Animate (bulk keyframes, smooth interpolation)
create_animation { "name": "animation.bear.walk", "loop": "loop", "length": 1.2 }
add_keyframes {
  "animation": "animation.bear.walk",
  "keyframes": [
    { "bone": "leg_fl", "channel": "rotation", "time": 0.0, "value": [28, 0, 0], "interpolation": "catmullrom" },
    { "bone": "leg_fl", "channel": "rotation", "time": 0.6, "value": [-28, 0, 0], "interpolation": "catmullrom" },
    { "bone": "leg_fl", "channel": "rotation", "time": 1.2, "value": [28, 0, 0], "interpolation": "catmullrom" }
  ]
}

// 7. Review from every side + audit, fix what you see, then repeat 4–7
screenshot_views { "views": ["isometric_right_front", "front", "left", "back"] }
check_model {}

// 8. Save / export
save_project   { "path": "D:/models/bear.bbmodel" }
export_project { "path": "D:/models/bear.geo.json" }

None of these tools knows what a hood or a scale is. They are extrusion, shells, arrays and chains.

// A hood: a shell with a real cavity, open at the face (north = -Z) and the neck (down)
add_hollow_volume {
  "bounds": { "from": [-5.5, 23, -5.5], "to": [5.5, 34, 5.5] },
  "wall_thickness": 1.5, "open_faces": ["north", "down"],
  "name": "hood", "parent": "head"
}

// A torn hem: 11 panels hanging off the cloak, tapering, jittered, staggered in depth so
// neighbours cannot z-fight
generate_array {
  "mode": "linear", "count": 11, "element_size": [2.2, 5, 1.2],
  "start": [-6, 7, 3.2], "end": [6, 7, 3.2],
  "anchor": "top", "jitter": [0.15, 0.4, 0], "size_decay": [0, -0.15, 0],
  "depth_stagger": 0.12, "rotation_range": { "min": [-4, 0, -9], "max": [4, 0, 9] },
  "seed": 12, "name_prefix": "hem", "parent": "body"
}

// A horn: 6 tapering segments, each bone bending 14° more than the last, animatable
extrude_chain {
  "segments": 6, "base_origin": [3.5, 32, -1], "segment_length": 2.4,
  "initial_size": [2.4, 2.4], "taper": 0.8, "curvature": [14, 0, -6],
  "direction": "up", "name": "horn", "side": "right", "parent": "head"
}

// A dragon wing: arm, forearm, 4 finger bones and a continuous membrane back to the body.
// Repeat with "side": "left" and base_origin x = -3 for the other wing.
add_wing {
  "side": "right", "base_origin": [3, 22, 2], "plane": "horizontal",
  "fingers": 4, "arm_length": 8, "forearm_length": 10, "finger_length": 18,
  "finger_spread": [0, 85], "parent": "chest"
}

// A blade: drawn as pixel art in the side plane, extruded into cubes
voxelize_matrix {
  "matrix": ["......####", ".....#####", "...#######", "..######..",
             ".#####....", "######....", "#####.....", "####......"],
  "palette": { "#": { "name": "blade", "depth": 1 } },
  "plane": "yz", "origin": [0, 18, -4], "merge_adjacent": true, "parent": "hand_right"
}

// The gate: is this actually a model yet?
audit_complexity { "target": "hero" }
// -> { "verdict": "acceptable", "ready_for_texturing": true, "cubes": 152,
//      "metrics": { "micro_pct": 34, "overlapping_pct": 61, "bone_depth": 6 }, "issues": [] }

FAQ

What is Blockbench MCP? It is a Model Context Protocol server plus a Blockbench plugin. Together they let an AI assistant control the Blockbench 3D editor: create and edit cubes, bones, textures and animations, and see the result through screenshots.

Which AI clients work with it? Any MCP client that supports stdio servers: Claude Code, Claude Desktop, Cursor, and others.

Which model formats are supported? Anything your Blockbench install supports: Minecraft Java Edition block/item models, Bedrock entities, GeckoLib animated models, generic models, and formats added by plugins (the AI can install those itself). list_formats shows what is available.

Can I use it to make Minecraft mobs and mod assets? Yes. The examples above are Minecraft-style mobs, and the GeckoLib workflow produces model, texture and animation files ready for a Forge / Fabric / NeoForge mod. export_model also writes glTF for Godot, Unity and Blender.

Does it work with the Blockbench web app? No. The bridge needs Node's net module, which only the desktop app exposes to plugins.

Troubleshooting

Blockbench shows "server stopped" and Start does nothing. The plugin needs the net module. When you click Start MCP Server, Blockbench shows a permission dialog; choose "Always allow for this plugin." If you denied it before, revoke and retry from the plugin's context menu, or restart Blockbench.

MCP tools fail with "Cannot reach Blockbench on 127.0.0.1:8787". Make sure Blockbench is open, the plugin is loaded and the server is running (green toast, or curl .../ping works). Confirm the port in the plugin settings matches BLOCKBENCH_MCP_PORT.

"Unknown format … the matching plugin must be installed." Plugin formats like GeckoLib's geckolib_model require install_plugin { "id": "geckolib" } first. Run list_formats to confirm the id appeared.

Console 404s about about.md / the plugin store. Harmless. Blockbench tries to fetch store metadata for the side-loaded plugin and gets a 404. It does not affect the bridge.

Security

  • The bridge binds to 127.0.0.1 only, so it is not reachable from your network.

  • Binding to localhost does not stop web pages in your browser from sending requests to it, so the bridge also refuses:

    • any request with an Origin header (browsers always send one on cross-origin requests; the MCP server does not);

    • any Host other than 127.0.0.1 / localhost / [::1] (blocks DNS rebinding);

    • POST bodies without Content-Type: application/json.

    It sends no CORS headers, so a page cannot read its responses either.

  • Trust boundary: the MCP client. Anything that controls the connected client, including a model steered by prompt injection from a file, web page or image it reads, can drive every tool.

  • Web access is limited to images. find_references / load_reference {url} run in the MCP server (not in Blockbench), force safe search, only fetch public http(s) addresses (localhost, LAN and link-local addresses are refused, including hostnames that resolve to them and every redirect hop), cap downloads at 15 MB and accept a file only if its bytes are a real PNG/JPEG/GIF/WebP/AVIF/BMP image.

  • execute_script runs unsandboxed JavaScript with Blockbench's full privileges. Only connect MCP clients you trust, and prefer the dedicated tools where possible. If you don't need it, turn off Settings ▸ General ▸ Allow execute_script. The command is then rejected inside Blockbench and every other tool keeps working.

Limitations

  • Desktop only. The bridge relies on Node modules (net), which the Blockbench web app does not expose to plugins.

  • One Blockbench window. The bridge talks to whichever project is currently active.

  • Formats with box UV only (e.g. Java entity / OptiFine) cannot give thin cubes their own pixels; pack_uv warns about them. Java block models only rotate one axis in 22.5° steps, so use 8 or 16 sides for round parts there.

  • find_references scrapes public image search pages; if an engine changes its markup it falls back to the next one.

Development

npm run dev     # tsc --watch
npm run build   # one-off compile to dist/
npm test        # build, then run the tool-catalogue tests against dist/

Issues and pull requests are welcome.


If Blockbench MCP saved you some modeling time, a ⭐ helps other people find it.

MIT © sosadly · Built on the Model Context Protocol · Not affiliated with Blockbench or GeckoLib.

Available Tools

46 tools
add_cubeA

Add a cube to the model. Coordinates are in Blockbench units. Cubes support rotation (degrees) and inflate (round/shrink without moving) — use them; flat axis-aligned boxes look robotic. For compound multi-axis angles, parent the cube to a rotated group instead. Prefer add_cubes to build many cubes at once. Returns the created cube with uuid and resolved face UVs (paint onto those with paint_faces).

Inputs: to (required), from (required), name, faces, +7 more.

add_cubesA

Create many cubes in one call — the efficient way to author a detailed model (aim for 20-50+ cubes for a creature, not 6-8). Pass cubes: an array where each item takes the same fields as add_cube ({name, from, to, origin, rotation, inflate, parent, box_uv, uv_offset, faces}). Build symmetric parts by emitting both the left side and its mirror (negate X, flip Y/Z rotation signs) in the same array. AVOID Z-FIGHTING: when cubes overlap, make one clearly penetrate the other (by >=0.1) and never align two faces to the exact same coordinate; stagger decorative pieces' depths. Returns all created cubes with their face UVs.

Inputs: cubes (required).

add_groupA

Add a group / bone to the outliner. Groups are the bones used for animation AND the way to apply free 3-axis rotation: a cube alone rotates cleanly on only one axis, so to pose a part at a compound angle, put it in a rotated group (nest groups for multi-axis angles). Set origin to the real joint so rotation pivots correctly. Returns the created group with its uuid.

Inputs: name, origin, parent, rotation.

add_groupsA

Create many bones/groups in one call — the fast way to lay out a whole skeleton. Pass groups: an array of {name, origin, rotation, parent}. A group's parent may reference another group created earlier in the SAME call by name, so you can build a nested, pre-posed bone hierarchy at once.

Inputs: groups (required).

add_keyframeA

Add a single keyframe to an animation for a given bone and channel.

Inputs: bone (required), time (required), value (required), animation (required), +2 more.

add_keyframesB

Add many keyframes at once — the efficient way to author a full animation. Pass an array of {bone, channel, time, value, interpolation}.

Inputs: animation (required), keyframes (required).

add_meshA

Create a non-cuboid MESH primitive so models aren't limited to axis-aligned boxes — crystals/gems/shards, pyramids, wedges, cones, cylinders, planes. Great for crystal cores, blades, horns, teeth, gems and stylised VFX. NOTE: meshes need a mesh-capable format (free/generic/bedrock); GeckoLib & Java export cubes only — for those build crystals from cubes rotated 45° instead.

Inputs: uv, from, name, size, +6 more.

add_planeA

Create a flat 2-sided plane (billboard) — the building block of pixel VFX (flames, energy sheets, slashes, motion trails) and for thin details (fins, leaves, paper). It is a zero-depth cube whose two large faces carry the texture; pair it with a VFX texture set to render_sides 'double'. crossed:true makes an X of two perpendicular planes for a volumetric particle look. Parent it to a bone to animate it.

Inputs: from (required), name, width, facing, +6 more.

apply_textureA

Apply a texture to all faces of an element (or all cubes if element omitted).

Inputs: texture (required), element.

check_modelA

Audit the model for problems that make results look broken: untextured faces (the 'gaps'), zero-area or out-of-bounds UVs, degenerate cube sizes, cubes not parented to a bone in animated formats, and Z-FIGHTING (coplanar_overlap — two faces on the same plane that flicker/clip, the 'two squares inside one another'). Run this after building and before/after texturing, then fix what it reports (for coplanar_overlap, nudge one cube by >=0.1 so the faces aren't coplanar). Returns a grouped issue list.

Inputs: no named parameters.

close_projectB

Close the currently open project.

Inputs: no named parameters.

create_animationA

Create an animation (requires a format that supports animation, e.g. GeckoLib animated_entity or Bedrock entity). Returns the animation uuid.

Inputs: loop, name, length.

create_textureB

Create a new texture. Either fill it with a solid color, or supply a full PNG via data_url. Returns the texture uuid.

Inputs: fill, name, width, height, +2 more.

create_vfx_textureA

Generate a pixelated VFX texture: a bright hot core fading to cool edges in quantized colour bands with jagged transparent edges — the look of pixel flames/energy/projectiles. With frames>1 it bakes a vertical FLIPBOOK and starts the animation player so the effect loops. Defaults to an additive/emissive render mode + 2-sided rendering so it glows on a plane. Apply it to add_plane planes (crossed/layered) and animate with bones. See get_guide topic 'vfx'.

Inputs: name, seed, style, width, +10 more.

delete_elementA

Delete a cube or group (and its children) from the model.

Inputs: element (required).

detail_cubesA

SMOOTH base texturing — the @volmur/Hytale look. Assigns the texture to every chosen face (no untextured 'gaps'), then per face bakes a soft vertical gradient in the region colour + gentle directional shading (top lighter, underside darker) + a SUBTLE low-contrast mottle, and finally a 3x3 box blur per UV island (the 'smooth brush'). Run pack_uv FIRST, then this right after create_texture, then paint_faces for crisp features. Avoids the dirty/noisy/grid look (no hard edge outline, low noise by default). Cubes named _core/_glow are filled bright (emissive read).

Inputs: base, blur, cubes, noise, +7 more.

edit_elementB

Edit an existing cube or group (rename, move, rotate, reparent, resize, inflate, visibility).

Inputs: element (required), to, from, origin, +5 more.

execute_scriptA

Run arbitrary JavaScript inside Blockbench's renderer for anything not covered by a dedicated tool. The code has access to all Blockbench globals (Project, Cube, Group, Texture, Animation, Undo, Canvas, Outliner, Format, Formats, ...) and receives a params object. Return a JSON-serializable value. Use sparingly; prefer dedicated tools.

Inputs: code (required), params.

export_projectC

Export the project through its format's codec (e.g. Java model JSON, Bedrock geometry, GeckoLib model). Provide path to write a file directly.

Inputs: path.

get_elementA

Get detailed info for one cube or group by uuid or name.

Inputs: element (required).

get_guideA

Return a playbook. Pass topic: 'modeling' (default — proportions, detail, rotation), 'texturing' (the smooth @volmur/Hytale look, no dirty noise), 'vfx' (pixelated flames/energy/projectiles/trails/auras via planes + emissive textures + animation), 'animation' (rigging, gaits, easing), or 'reference' (how to ACTUALLY match a reference image instead of 'almost'). READ the relevant topic BEFORE building/texturing/animating — it dramatically improves results.

Inputs: topic.

get_statusA

Get the current Blockbench state: open project, format, counts of cubes/groups/textures/animations, and edit mode. Call this first to understand the workspace. If you are about to BUILD or TEXTURE a model, call get_guide first.

Inputs: no named parameters.

get_textureA

Read a texture back as an image so you can inspect what it currently looks like. Returns the PNG inline.

Inputs: texture (required).

import_textureB

Import a texture from an image file on disk (desktop only).

Inputs: path (required), name.

install_pluginA

Install a Blockbench plugin from the store (by id, e.g. 'geckolib' for GeckoLib Models & Animations), or from a url, or a local path. Needed before using plugin-specific formats like GeckoLib's 'geckolib_model'.

Inputs: id, url, path.

list_animationsA

List all animations and their animated bones.

Inputs: no named parameters.

list_formatsA

List all model formats available in this Blockbench install (e.g. free, java_block, bedrock, and any added by plugins such as GeckoLib's animated_entity). Use the returned id with new_project.

Inputs: no named parameters.

list_outlinerA

Return the full outliner tree (groups/bones and their nested cubes) with uuids, origins and rotations.

Inputs: no named parameters.

list_pluginsA

List Blockbench plugins (installed and available in the store). Filter with query or installed_only.

Inputs: query, installed_only.

list_texturesA

List all textures in the project.

Inputs: no named parameters.

load_projectA

Load a .bbmodel project file from disk (desktop only).

Inputs: path (required).

mirror_elementA

Mirror a cube or group (with its children) across an axis about a pivot — build one side of a symmetric model, then mirror it. Flips geometry and the off-axis rotation signs, and renames left<->right. Returns the created clones.

Inputs: axis, pivot, element, elements.

new_projectA

Create a new project from the start screen, choosing a format. This is the entry point for any new model.

Inputs: format (required), name, geometry_name, texture_width, +1 more.

pack_uvA

Shelf-pack the box UVs so every cube gets its own region of the texture. REQUIRED before texturing a box_uv model (GeckoLib/Bedrock): newly created cubes all sit at uv_offset [0,0] and otherwise paint onto the same pixels. Re-run after adding or resizing cubes. Auto-grows the texture (preserving paint) if the layout overflows.

Inputs: cubes, padding, auto_resize.

paint_facesA

Paint features onto specific cube faces using coordinates RELATIVE to each face (so [0,0] is that face's top-left corner) — no manual UV math, which is what usually causes misplaced/garbled texture. Use it for eyes, nostrils, mouths, claws, fur tufts, stripes, scars, bandages, armour trim, etc. Either pass one {cube, face, base?, ops?} or a faces array of them.

Inputs: ops, base, cube, face, +2 more.

paint_textureA

Paint directly on a texture with absolute pixel coordinates. Use this for whole-sheet work; for painting onto a specific cube face, paint_faces (face-relative coords) is usually easier. Ops run in order on the canvas (origin top-left, y down).

Inputs: ops (required), texture (required), edit_name.

remove_animationC

Delete an animation from the project.

Inputs: animation (required).

resize_textureC

Resize a texture's bitmap to new dimensions (nearest-neighbour).

Inputs: width (required), height (required), texture (required).

save_projectA

Save the open project as a .bbmodel. Provide path to save to a specific file (desktop), otherwise Blockbench's save flow is used.

Inputs: path.

screenshotA

Capture the current 3D preview and return it as an image so you can visually inspect the model and iterate. Optionally specify width/height.

Inputs: width, height.

screenshot_viewsB

Capture several camera angles in ONE call and return them all as images, so you can see the whole model and catch problems (gaps, wrong rotations, missing detail, asymmetry) from every side. Defaults to iso/front/left/back. This is the main way to review and iterate — do it after each modeling/texturing pass, not just once.

Inputs: views, width, height.

set_camera_angleA

Position the preview camera, by named preset and/or explicit camera position & target.

Inputs: angle, preset, target, position.

set_cube_uvA

Set UV mapping and/or per-face texture on a cube's faces.

Inputs: cube (required), faces (required).

set_project_metaC

Update the open project's name, geometry name, or texture resolution.

Inputs: name, geometry_name, texture_width, texture_height.

set_texture_render_modeA

Set how a texture renders: render_mode ('default' | 'emissive' = full-bright, ignores light | 'additive' = bright pixels add light & dark vanishes, best for fire/energy on planes | 'layered' | 'normal' | 'height' | 'mer'), render_sides ('auto' | 'front' | 'double' for 2-sided planes), flipbook frame timing, and particle flag. Use this to make VFX glow and to show planes from both sides.

Inputs: texture (required), animate, particle, frame_time, +4 more.

uninstall_pluginB

Uninstall an installed Blockbench plugin by id.

Inputs: id (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.7/5.0

    Scored across 46 tools

    Disambiguation5/5

    Each tool has a distinct purpose, with clear separation between single-item and batch operations (e.g., add_cube vs add_cubes). Texturing tools are well-differentiated (paint_faces vs paint_texture vs detail_cubes), and even execute_script is reserved for edge cases. No two tools appear to do the same thing.

    Naming Consistency5/5

    Tool names follow a consistent verb_noun pattern using snake_case (e.g., add_cube, create_animation, list_formats). Deviations like new_project and check_model are minor and still intuitive. The naming is predictable and easy to navigate.

    Tool Count4/5

    At 46 tools, the set is larger than typical MCP servers but justified by the complexity of 3D modeling. Each tool addresses a specific, non-trivial operation (e.g., pack_uv, detail_cubes). A few could be merged (e.g., add_cube/add_cubes) but the batch versions are genuinely needed.

    Completeness5/5

    The toolset covers the full modeling pipeline: project creation, element construction (cubes, groups, meshes, planes), texturing (creation, UV packing, painting, style application), animation, inspection/auditing (check_model, screenshot_views), and plugin management. There are no major gaps; even edge cases are addressed via execute_script.

    Maintenance

    ActivityMaintained
    ResponsivenessLimited data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI assistants to control Blockbench for Minecraft 3D modeling, including project creation, cube placement, UV layout, and texture painting.
      GPL 3.0
    • A
      license
      Not graded
      quality
      D
      maintenance
      Connects Claude AI to Blockbench through the Model Context Protocol, enabling AI-assisted 3D modeling, texture creation, and block model manipulation. It supports creating, modifying, and deleting models, applying textures and materials, and tracking commands in real time.
      MIT