Edge Veda MCP Server for Claude Code

An MCP (Model Context Protocol) server that lets Claude Code automate Edge Veda Flutter project setup end-to-end. Instead of following a 20-step quickstart manually, just tell Claude Code what you want to build and the MCP tools handle environment checks, project scaffolding, model selection, and device deployment.

What This Does

Reduces the 3-hour beginner journey to ~15 minutes by providing 6 tools that Claude Code calls automatically:

1. Check environment -- verify Flutter, Xcode, CocoaPods, iOS device
2. List models -- device-aware recommendations with memory fit estimates
3. Create project -- scaffold Flutter project with Edge Veda configured
4. Download model -- fetch the right GGUF file for the use case
5. Add capability -- inject code scaffolding for chat, vision, STT, TTS, image, RAG
6. Run -- build and deploy to connected iOS device

Prerequisites

• Node.js >= 18
• Flutter SDK
• Xcode (for iOS development)
• CocoaPods

Installation

One-liner (npx)

No install needed -- npx fetches and runs the latest version:

npx @edge-veda/mcp-server

Add to Claude Code

claude mcp add edge-veda -- npx @edge-veda/mcp-server

Add to Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "edge-veda": {
      "command": "npx",
      "args": ["@edge-veda/mcp-server"]
    }
  }
}

From Source (contributors)

git clone https://github.com/ramanujammv1988/edge-veda.git
cd edge-veda/tools/mcp-server
npm install
npm run build

Available Tools

| Tool | Description | Key Parameters | |------|-------------|---------------| | edge_veda_check_environment | Verify dev prerequisites are installed | (none) | | edge_veda_list_models | List models with device-aware recommendations | use_case?, show_all? | | edge_veda_create_project | Scaffold a Flutter project with Edge Veda | project_name, path?, model_id? | | edge_veda_download_model | Download a model GGUF file | model_id, project_path | | edge_veda_add_capability | Add capability code scaffolding | capability, project_path | | edge_veda_run | Build and deploy to iOS device | project_path, mode?, device_id? |

Tool Details

edge_veda_check_environment

No parameters. Checks for Flutter SDK, Xcode, CocoaPods, and connected iOS devices. Returns a pass/fail report.

edge_veda_list_models

• use_case (optional): Filter by chat, vision, stt, tts, image, embedding, tool-calling
• show_all (optional): Include projector files and large desktop models

Returns a table of models with size, estimated memory usage, device fit status, and the recommended pick.

edge_veda_create_project

• project_name (required): Flutter project name (lowercase_with_underscores)
• path (optional): Parent directory (defaults to cwd)
• model_id (optional): Model ID for the boilerplate (defaults to llama-3.2-1b-instruct-q4). Valid IDs: llama-3.2-1b-instruct-q4, phi-3.5-mini-instruct-q4, gemma-2-2b-instruct-q4, qwen3-0.6b-q4, tinyllama-1.1b-chat-q4.

Runs flutter create, adds edge_veda: ^2.4.2 to pubspec.yaml, runs flutter pub get, sets iOS 13.0 minimum, configures use_modular_headers! for FFI compatibility, runs pod install, and writes a working boilerplate main.dart with model download + streaming inference.

edge_veda_download_model

• model_id (required): Model ID from edge_veda_list_models (e.g., llama-3.2-1b-instruct-q4)
• project_path (required): Path to the Flutter project

Downloads the model file to {project}/models/ and adds models/ to .gitignore. This is optional -- the app auto-downloads models at runtime via ModelManager.downloadModel().

edge_veda_add_capability

• capability (required): One of chat, vision, stt, tts, image, rag
• project_path (required): Path to the Flutter project

Creates a new screen file (lib/{capability}_screen.dart) with complete working code, runs flutter pub get, and provides main.dart wiring instructions. Lists required models that need downloading. Adds extra dependencies to pubspec.yaml if needed.

edge_veda_run

• project_path (required): Path to the Flutter project
• mode (optional): debug, profile, or release (defaults to release)
• device_id (optional): Device UDID (auto-detects connected device)

Builds and deploys the app. Defaults to release mode for optimal inference speed (~42 tok/s vs ~5 tok/s in debug).

Example Workflow

Here is a natural language conversation showing how Claude Code uses the tools:

You: "I want to build an iOS app with on-device chat"

Claude Code:

1. Calls edge_veda_check_environment -- verifies Flutter, Xcode, CocoaPods are installed
2. Calls edge_veda_create_project with project_name: "my_chat_app" -- scaffolds the project
3. Calls edge_veda_list_models with use_case: "chat" -- shows Llama 3.2 1B as recommended
4. Calls edge_veda_run with project_path: "./my_chat_app" -- deploys to your iPhone

The boilerplate main.dart auto-downloads the model on first launch, so the app is ready to chat.

You: "Add vision capability too"

Claude Code:

1. Calls edge_veda_add_capability with capability: "vision" -- creates lib/vision_screen.dart, runs flutter pub get automatically, and provides wiring instructions
2. Calls edge_veda_download_model with model_id: "smolvlm2-500m-video-instruct-q8" -- pre-downloads the model to {project}/models/
3. Calls edge_veda_run -- rebuilds and deploys the updated app

Supported Models

The MCP server includes the complete Edge Veda model registry:

Text/Chat: Llama 3.1 8B, Mistral Nemo 12B, Llama 3.2 1B, Phi 3.5 Mini, Gemma 2 2B, TinyLlama 1.1B, Qwen3 0.6B

Vision: SmolVLM2 500M, LLaVA 1.6 Mistral 7B, Qwen2-VL 7B (each with mmproj)

Speech-to-Text: Whisper Tiny/Base (English), Whisper Small/Medium/Large v3 (Multilingual)

Embedding: All MiniLM L6 v2, Nomic Embed Text, mxbai-embed-large

Image Generation: SD v2.1 Turbo, SDXL Turbo, FLUX.1 Schnell

Troubleshooting

Server does not start

Ensure Node.js >= 18 is installed:

node --version

Test the server directly:

npx @edge-veda/mcp-server

It should hang waiting for JSON-RPC input on stdin. Press Ctrl+C to exit.

Tools not appearing in Claude Code

1. Verify the MCP server is registered: claude mcp list
2. Re-add if needed: claude mcp add edge-veda -- npx @edge-veda/mcp-server
3. Restart Claude Code after changing MCP configuration

flutter create fails

Ensure Flutter is on your PATH:

flutter doctor

pod install fails

Clear CocoaPods cache and retry:

cd ios && pod deintegrate && pod install

Architecture

tools/mcp-server/
  src/
    index.ts              # MCP server entry point (stdio transport)
    utils.ts              # exec helpers
    model-registry.ts     # TypeScript mirror of Dart ModelRegistry
    device-profile.ts     # Device tier detection + memory estimation
    tools/
      check-environment.ts
      create-project.ts
      list-models.ts
      download-model.ts
      add-capability.ts
      run.ts
  build/                  # Compiled JS output
  test/
    smoke.test.ts         # MCP protocol smoke test

The server uses the official MCP SDK (@modelcontextprotocol/sdk) with stdio transport. All communication happens over stdin/stdout using JSON-RPC 2.0. Diagnostic output goes to stderr only.