MCP Runware Server

✨ Vercel-ready with Fluid compute | Serverless MCP deployment with streaming support

An MCP (Model Context Protocol) server for AI image generation through Runware's API. Generate images from text using FLUX.1, FLUX.2, HiDream, and more with full control over dimensions, steps, schedulers, and other parameters.

✨ Fully compatible with Vercel with Fluid compute - Deploy as a serverless MCP endpoint with streaming support.

Quick Start

# Install
npm install

# Configure
cp .env.example .env
# Edit .env and add your RUNWARE_API_KEY

# Build and run
npm run build
npm start

Prerequisites: Node.js 22+, Runware API key (get one)

Environment Variables

Create a .env file in the project root:

# Required - Your Runware API key
RUNWARE_API_KEY=rwk_abc123xyz456

# Optional - Transport mode (stdio|http)
# - stdio: Local Claude Desktop (default)
# - http: HTTP server for remote connections
MCP_TRANSPORT=stdio

# Optional - HTTP server configuration (http mode only)
PORT=3000
HOST=127.0.0.1

# Optional - API key authentication (http mode)
# Comma-separated list for multiple keys
MCP_API_KEYS=secret-key-1,secret-key-2

Variable Details

RUNWARE_API_KEY (required): Get from Runware Dashboard → API Keys. Used to authenticate with Runware's image generation API.
MCP_TRANSPORT: Controls how the server communicates with clients:
- stdio - Standard input/output for local Claude Desktop (default)
- http - HTTP server with /mcp endpoint for remote access
PORT / HOST: HTTP server binding (default: 3000 / 127.0.0.1). Only used when MCP_TRANSPORT=http.
MCP_API_KEYS: Security for HTTP/Vercel modes. Without this, the server accepts all requests (dev mode). With keys, clients must send x-api-key header. Health checks (/health) bypass authentication.

Usage

Claude Desktop (Local)

Add to your configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

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

Restart Claude Desktop. The server loads environment variables from .env automatically.

HTTP Server (Remote Access)

# Start HTTP server
npm run start:http

# Test endpoints
curl http://127.0.0.1:3000/health
curl -X POST http://127.0.0.1:3000/mcp \
  -H "x-api-key: your-key" \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize",...}'

Connect from MCP client:

import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";

const transport = new StreamableHTTPClientTransport(
  new URL("http://127.0.0.1:3000/mcp"),
  { headers: { "x-api-key": "your-key" } },
);

Vercel Deployment (Serverless)

✨ Fully compatible with Vercel Fluid compute for streaming MCP responses.

Connect repo to Vercel - Import your Git repository
Set environment variables in Vercel dashboard:
- RUNWARE_API_KEY - Your Runware API key (required)
- MCP_API_KEYS - Authentication keys, comma-separated (recommended)
Deploy - Vercel auto-detects configuration from vercel.json with Fluid enabled

Your MCP server: https://your-project.vercel.app/mcp

Connect with same client code as HTTP mode using your Vercel URL. The server leverages Vercel's experimental Fluid compute to support streaming responses required by MCP.

Important Notes:

Session Management: Vercel's serverless functions are stateless. Each session creates its own MCP server instance. If a client receives a "Session not found" error, it should reinitialize by sending a new initialize request with the same mcp-session-id header to recreate the session.
Model Data: The build process automatically copies model data from src/data/ to dist/data/. After updating models with npm run fetch-models, rebuild before deploying.
Cold Starts: Initial requests may be slower due to serverless cold starts. Subsequent requests within the same instance are faster.

Testing

MCP Inspector (Recommended)

Visual testing with MCP Inspector:

npm run build
npx @modelcontextprotocol/inspector node dist/index.js

OR using HTTP+Streamable

npm run build
npm run start:http
npm run inspect

Opens browser UI at http://localhost:6274 to test tools interactively. If you see SSE errors, check .env has valid RUNWARE_API_KEY.

Command Line

# List tools
echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | node dist/index.js

# Generate image
cat <<EOF | node dist/index.js
{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"generate_image","arguments":{"prompt":"Mountain sunset","width":1024,"height":768}}}
EOF

Available Tools

generate_image

Generate an image from a text prompt.

Parameters:

prompt (required): Text description of the image (2-3000 characters)
model (optional): Model ID (default: runware:101@1 - FLUX.1 Dev)
negativePrompt (optional): What to avoid in the image
width (optional): Width in pixels, divisible by 64 (128-2048, default: 1024)
height (optional): Height in pixels, divisible by 64 (128-2048, default: 1024)
steps (optional): Generation steps (1-100, default: 20)
cfgScale (optional): Guidance scale (0-50, default: 7)
scheduler (optional): Sampling scheduler (e.g., "DPM++ 2M Karras", "Euler A")
seed (optional): Seed for reproducible results
numberResults (optional): Number of images to generate (default: 1)
includeCost (optional): Include cost in response (default: true)

Example:

{
  "prompt": "A futuristic city at sunset with flying cars and neon lights",
  "model": "runware:101@1",
  "width": 1024,
  "height": 768,
  "steps": 30,
  "cfgScale": 7
}

get_models

Get a comprehensive list of AI models available on Runware with their AIR identifiers, pricing, and descriptions. Models are automatically sorted by price (cheapest first).

No parameters required

Returns: A merged, deduplicated list of 44+ models from popular and curated collections.

Features:

Pricing information (USD per image/generation)
Model configurations and discounts
AIR identifiers for direct usage with generate_image tool
Combines manually curated popular models and "Best for Text on Images" collection
Sorted by price with cheapest models first

Example Response:

[
  {
    "name": "Stable Diffusion XL Lightning",
    "air": "civitai:133005@782002",
    "description": "Fast SDXL with 4-step generation",
    "price_usd": 0.001,
    "configuration": "4-step",
    "discount": "80% cheaper"
  },
  {
    "name": "FLUX.1 Schnell",
    "air": "runware:100@1",
    "description": "Ultra-fast FLUX variant",
    "price_usd": 0.003,
    "configuration": "4-step",
    "discount": "70% cheaper"
  }
  // ... more models
]

Model Data Management:

Model information stored in src/data/popular_models.json and src/data/best_models.json
Pricing data in src/data/pricing.json
Update models: npm run fetch-models (requires RUNWARE_API_KEY in .env)
Build process automatically copies data from src/data/ to dist/data/
Rebuild and redeploy after updating model data

Returns:

A formatted list of all available models including:

Model name and AIR identifier
Pricing information (USD per generation)
Configuration details (resolution, steps, etc.)
Discount information where applicable
Model category and tags

The tool merges data from both popular models and specialized collections, providing up to 44+ unique models sorted by cost-effectiveness.

Example output:

FLUX.2 [klein] 4B (runware:400@4) - $0.0006 (1024x1024) [Save 40%]
Z-Image-Turbo (runware:z-image@turbo) - $0.0006 (1024x1024 · 4 steps)
FLUX.2 [klein] 9B (runware:400@2) - $0.00078 (1024x1024 · 4 steps) [Save 87%]
Qwen-Image-2512 (alibaba:qwen-image@2512) - $0.0051 (1024x1024) [Save 74%]

All pricing and model data is automatically updated when running npm run fetch-models.

Project Structure

mcp-runware/
├── src/
│   ├── index.ts              # Main entry point
│   ├── transports.ts         # Transport setup (stdio, HTTP)
│   ├── utils.ts              # Shared utilities (server factory, auth, transport factory)
│   ├── tools.ts              # MCP tool definitions
│   ├── runware-client.ts     # Runware SDK wrapper
│   └── data/                 # Model and pricing data
│       ├── popular_models.json    # Popular models with AIR & pricing
│       ├── best_models.json       # Best text-on-images models
│       └── pricing.json           # Pricing data source
├── scripts/
│   └── fetch_curated_models.py    # Model data fetcher/enricher
├── api/
│   └── index.ts              # Vercel serverless function
├── dist/                     # Compiled JavaScript (generated)
├── vercel.json               # Vercel deployment configuration
├── .env                      # Environment variables (create from .env.example)
├── .env.example              # Environment variables template
├── package.json              # Dependencies and scripts
├── tsconfig.json             # TypeScript configuration
└── README.md                 # This file

Development

Watch Mode

For active development with auto-recompilation:

npm run watch

In another terminal, run the server:

node dist/index.js

Build

Compile TypeScript to JavaScript:

npm run build

Run

Start the MCP server:

npm start

Or build and run in one command:with pricing information, automatically scraped and enriched from Runware. To refresh these lists:

Install Python dependencies (one-time setup):
```
pip3 install requests python-dotenv
```
Run the fetch script:
```
npm run fetch-models
```

This will:

Scrape model IDs and names from Runware collection pages
Fetch AIR identifiers and metadata from the Runware API
Merge pricing data from src/data/pricing.json
Filter out models without AIR identifiers
Save enriched data to JSON files in src/data/

Data sources:

Popular Models → src/data/popular_models.json (manually curated from https://runware.ai/models)
Best Models → src/data/best_models.json (scraped from https://runware.ai/collections/best-for-text-on-images)
Pricing Data → src/data/pricing.json (extracted from https://runware.ai/pricing)

The get_models tool automatically merges these sources and sorts by price.

To modify the model collections, edit the POPULAR_MODELS and SCRAPE_COLLECTIONS arrays API

Save enriched data to JSON files in src/data/

Collections fetched:

Popular Models → src/data/popular_models.json (from https://runware.ai/models)
Best Models → src/data/best_models.json (from https://runware.ai/collections/best-for-text-on-images)

To add more collections, edit the COLLECTIONS array in scripts/fetch_curated_models.py.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

MCP Runware Server

Quick Start

Environment Variables

Variable Details

Usage

Claude Desktop (Local)

HTTP Server (Remote Access)

Vercel Deployment (Serverless)

Testing

MCP Inspector (Recommended)

Command Line

Available Tools

generate_image

get_models

Project Structure

Development

Watch Mode

Build

Run

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Name		Name	Last commit message	Last commit date
Latest commit History 8 Commits
api		api
scripts		scripts
src		src
.env.example		.env.example
.gitignore		.gitignore
.vercelignore		.vercelignore
README.md		README.md
claude_desktop_config.example.json		claude_desktop_config.example.json
package-lock.json		package-lock.json
package.json		package.json
test-simple.ts		test-simple.ts
tsconfig.json		tsconfig.json
vercel.json		vercel.json

Folders and files

Latest commit

History

Repository files navigation

MCP Runware Server

Quick Start

Environment Variables

Variable Details

Usage

Claude Desktop (Local)

HTTP Server (Remote Access)

Vercel Deployment (Serverless)

Testing

MCP Inspector (Recommended)

Command Line

Available Tools

generate_image

get_models

Project Structure

Development

Watch Mode

Build

Run

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages