logoMCP

A local MCP server that generates complete brand design systems using AI. Give it a brand brief, get back logo concepts, color palettes, typography, and a full asset library — all running locally on Apple Silicon.

Built for LLM-based marketing teams: Claude or other agents submit briefs, visually review generated logos, iterate on brand specs, and export production-ready assets.

What It Does

1. Brand Brief → You describe the brand (name, industry, audience, mood)
2. AI Brand Specs → Local text LLM generates creative direction, tagline, color palette, typography
3. Logo Concepts → Local image model generates 3 logo mark concepts at 1024×1024
4. Template Rendering → Selecting a logo triggers automatic generation of all brand assets:
   • Logo variants (mark, primary horizontal, stacked, wordmark, favicon)
   • Each in light/dark/transparent backgrounds
   • PNG (1x + 2x), JPEG, and SVG formats
   • Favicon at 16, 32, 48, and 192px

~3 minutes from brief to 47 production-ready assets. No cloud APIs, no subscriptions, everything on your Mac.

Requirements

• macOS with Apple Silicon (M1 or later)
• 24GB+ RAM (32GB recommended)
• Python 3.12+ via uv
• HuggingFace account with access to FLUX.1-schnell (free, gated model — accept terms then hf auth login)

Installation

# Clone
git clone https://github.com/gofastercloud/logoMCP.git
cd logoMCP

# Install Python dependencies
cd backend
uv sync

# Authenticate with HuggingFace (required for Flux model access)
hf auth login
# Or: cd backend && uv run python -c "from huggingface_hub import login; login()"

# Run tests to verify everything works
uv run pytest

Models are downloaded automatically on first use (~12GB total for Qwen3 8B + Flux.1-schnell).

Using with Claude Code

Add logoMCP as an MCP server in your Claude Code settings:

Option 1: Project-level — add to .claude/settings.json in any project:

{
  "mcpServers": {
    "logomcp": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/logoMCP/backend", "python", "-m", "logogen.server"],
      "env": {}
    }
  }
}

Option 2: User-level — add to ~/.claude/settings.json to make it available in all projects:

{
  "mcpServers": {
    "logomcp": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/logoMCP/backend", "python", "-m", "logogen.server"],
      "env": {}
    }
  }
}

Replace /path/to/logoMCP with the actual path where you cloned the repo.

Using with Claude Code Cowork

When running Claude Code in cowork/team mode, any team member with access to the MCP server can call logoMCP tools. Configure the MCP server at the user level (Option 2 above) so all spawned agents inherit access.

Example cowork workflow — a marketing team lead agent can:

1. create_brand("Solaris", "Solar Energy", "Eco-conscious homeowners", ["sustainable", "bright"])
2. update_brand_specs(brand_id, colors=[...], typography={...})  # optional pre-set preferences
3. generate_logos(brand_id)  # generates specs + 3 logo concepts
4. [visually review the 3 returned concept images]
5. select_logo(brand_id, concept_index=1)  # triggers template rendering
6. get_brand_assets(brand_id)  # browse all rendered assets
7. get_asset(brand_id, "logo-primary", "dark", "png")  # inspect specific asset
8. export_brand(brand_id)  # download ZIP

MCP Tools

Tool	Description
create_brand	Create a brand from a brief (name, industry, audience, mood keywords)
list_brands	List all brands with status
get_brand	Get full brand details (specs, concepts, assets)
delete_brand	Delete a brand and all generated assets
update_brand_specs	Set or override colors, fonts, tagline, creative direction
generate_logos	Run AI pipeline: text LLM → image model → 3 logo concepts
get_logo_concepts	Review logo concept images
select_logo	Pick a concept → triggers template rendering of all assets
get_logo_prompts	View current image generation prompts
update_logo_prompts	Revise prompts (with safety guardrails enforced)
get_brand_assets	Browse rendered assets with thumbnails
get_asset	Get a specific asset image in any format/variant/scale

Architecture

Brand Brief → Text LLM (Qwen3 8B) → Image Model (Flux.1-schnell) → Template Engine → Assets
                  ↓                         ↓                            ↓
           Creative direction        3 logo concepts           47 production files
           Color palette             1024×1024 each            PNG/JPEG/SVG
           Typography                                          Light/Dark/Transparent
           Tagline                                             1x + 2x resolution

Memory management: Models load sequentially — text model generates all specs, unloads, then image model generates logos, unloads. Template rendering is pure CPU (Pillow). This keeps peak memory under 16GB, fitting comfortably on 24GB machines.

Template system: Each asset type is a Python class inheriting from BaseTemplate. Adding a new asset = writing a render() method. The engine handles all variant/format/scale permutations automatically.

Configuration

Copy .env.example to .env and customize:

# Model selection
TEXT_MODEL_ID=mlx-community/Qwen3-8B-4bit    # Text LLM
IMAGE_MODEL_NAME=schnell                       # Flux variant (schnell or dev)
IMAGE_MODEL_QUANTIZE=4                         # Quantization (4 or 8 bit)

# Generation parameters
LOGO_STEPS=4          # Inference steps (schnell: 2-4, dev: 20-25)
LOGO_WIDTH=1024       # Logo dimensions (must be multiples of 32)
LOGO_HEIGHT=1024

Use IMAGE_MODEL_NAME=dev for higher quality logos (slower, needs 32GB+ RAM).

Development

cd backend

# Run tests
uv run pytest

# Run tests verbose
uv run pytest -v

# Run E2E test with real models (slow, generates actual logos)
uv run python test_e2e.py

# Run MCP server directly (for debugging)
uv run python -m logogen.server

Dependency Licenses

All dependencies use permissive licenses compatible with MIT:

Dependency	License	Purpose
mlx / mlx-lm	MIT	Apple Silicon ML framework + LLM inference
mflux	MIT	Flux image generation on MLX
Pillow	HPND	Image compositing and template rendering
drawsvg	MIT	SVG generation
vtracer	MIT	Bitmap to SVG tracing
pydantic	MIT	Data validation and schemas
aiosqlite	MIT	Async SQLite
mcp	MIT	MCP server SDK
python-dotenv	BSD-3-Clause	Environment variable loading

License

MIT — see LICENSE.