VisionAgent MCP Server

> Beta – v0.1

> This project is early access and subject to breaking changes until v1.0.

VisionAgent MCP Server v0.1 - Overview

Modern LLM “agents” call external tools through the **Model Context Protocol (MCP). VisionAgent MCP** is a lightweight, side-car MCP server that runs locally on STDIN/STDOUT, translating each tool call from an MCP-compatible client (Claude Desktop, Cursor, Cline, etc.) into an authenticated HTTPS request to Landing AI’s VisionAgent REST APIs. The response JSON, plus any images or masks, is streamed back to the model so that you can issue natural-language computer-vision and document-analysis commands from your editor without writing custom REST code or loading an extra SDK.

📸 Demo

https://github.com/user-attachments/assets/2017fa01-0e7f-411c-a417-9f79562627b7

🧰 Supported Use Cases (v0.1)

> Run `npm run generate-tools` whenever VisionAgent releases new endpoints. The script fetches the latest OpenAPI spec and regenerates the local tool map automatically.

🗺 Table of Contents

1. Quick Start

2. Configuration

3. Example Prompts

4. Architecture & Flow

5. Developer Guide

6. Troubleshooting

7. Contributing

8. Security & Privacy

🚀 Quick Start

Get Your VisionAgent API Key

If you do not have a VisionAgent API key, create an account and obtain your API key.

# 1  Install
npm install -g vision-tools-mcp

# 2  Configure your MCP client with the following settings:
{
  "mcpServers": {
    "VisionAgent": {
      "command": "npx",
      "args": ["vision-tools-mcp"],
      "env": {
        "VISION_AGENT_API_KEY": "",
        "OUTPUT_DIRECTORY": "/path/to/output/directory",
        "IMAGE_DISPLAY_ENABLED": "true" # or false, see below
      }
    }
  }
}

3. Open your MCP-aware client.

4. Download *street.png* (from the assets folder in this directory, or you can choose any test image).

5. Paste the prompt below (or any prompt):

Detect all traffic lights in /path/to/mcp/vision-agent-mcp/assets/street.png

If your client supports inline resources, you’ll see bounding-box overlays; otherwise, the PNG is saved to your output directory, and the chat shows its path.

Prerequisites

⚙️ Configuration

Sample MCP client entry (`.mcp.json` for VS Code / Cursor)

{
  "mcpServers": {
    "VisionAgent": {
      "command": "npx",
      "args": ["vision-tools-mcp"],
      "env": {
        "VISION_AGENT_API_KEY": "912jkefief09jfjkMfoklwOWdp9293jefklwfweLQWO9jfjkMfoklwDK",
        "OUTPUT_DIRECTORY": "/Users/me/documents/mcp/test",
        "IMAGE_DISPLAY_ENABLED": "false"
      }
    }
  }
}

For MCP clients without image display capabilities, like Cursor, set IMAGE_DISPLAY_ENABLED to False. For MCP clients with image display capabilities, like Claude Desktop, set IMAGE_DISPLAY_ENABLED to true to visualize tool outputs. Generally, MCP clients that support resources (see this list: https://modelcontextprotocol.io/clients) will support image display.

💡 Example Prompts

🏗 Architecture & Flow

┌────────────────────┐ 1. human prompt            ┌───────────────────┐
│ MCP-capable client │───────────────────────────▶│  VisionAgent MCP │
│  (Cursor, Claude)  │                            │   (this repo)     │
└────────────────────┘                            └─────────▲─────────┘
            ▲  6. rendered PNG / JSON                     │ 2. JSON tool call
            │                                             │
            │ 5. preview path / data         3. HTTPS     │
            │                                             ▼
       local disk  ◀──────────┐                Landing AI VisionAgent
                               └──────────────  Cloud APIs
                                           4. JSON / media blob

1. Prompt → tool-call The client converts your natural-language prompt into a structured MCP call.

2. Validation The server validates args with Zod schemas derived from the live OpenAPI spec.

3. Forward An authenticated Axios request hits the VisionAgent endpoint.

4. Response JSON + any base64 media are returned.

5. Visualization If enabled, masks / boxes / depth maps are rendered to files.

6. Return to chat The MCP client receives data + file paths (or inline previews).

🧑‍💻 Developer Guide

Here’s how to dive into the code, add new endpoints, or troubleshoot issues.

Installation & Build

1. Clone the repository:

git clone https://github.com/landing-ai/vision-agent-mcp.git

2. Navigate into the project directory:

cd vision-agent-mcp

3. Install dependencies:

npm install

4. Build the project:

npm run build

Environment Variables

• `VISION_AGENT_API_KEY` - Required API key for VisionAgent authentication
• `OUTPUT_DIRECTORY` - Optional directory for saving processed outputs (supports relative and absolute paths)
• `IMAGE_DISPLAY_ENABLED` - Set to `"true"` to enable image visualization features

Client Configuration

After building, configure your MCP client with the following settings:

{
  "mcpServers": {
    "VisionAgent": {
      "command": "node",
      "args": [
        "/path/to/build/index.js"
      ],
      "env": {
        "VISION_AGENT_API_KEY": "",
        "OUTPUT_DIRECTORY": "../../output",
        "IMAGE_DISPLAY_ENABLED": "true"
      }
    }
  }
}

> Note: Replace `/path/to/build/index.js` with the actual path to your built `index.js` file, and set your environment variables as needed. For MCP clients without image display capabilities, like Cursor, set IMAGE_DISPLAY_ENABLED to False. For MCP clients with image display capabilities, like Claude Desktop, set IMAGE_DISPLAY_ENABLED to true to visualize tool outputs. Generally, MCP clients that support resources (see this list: https://modelcontextprotocol.io/clients) will support image display.

📑 Scripts & Commands

> Pro Tip: If you modify any files under `src/` or want to pick up new endpoints from VisionAgent, run `npm run build:all` to recompile + regenerate tool definitions.

📂 Project Layout

vision-agent-mcp/
├── .eslintrc.json              # ESLint config (optional)
├── .gitignore                  # Ignore node_modules, build/, .env, etc.
├── jest.config.js              # Placeholder for future unit tests
├── mcp-va.md                   # Draft docs (incomplete)
├── package.json                # npm metadata, scripts, dependencies
├── package-lock.json           # Lockfile
├── tsconfig.json               # TypeScript compiler config
├── .env                        # Your environment variables (not committed)
│
├── src/                        # TypeScript source code
│   ├── generateTools.ts        # Dev script: fetch OpenAPI → generate MCP tool definitions (Zod schemas)
│   ├── index.ts                # Entry point: load .env, start MCP server, handle signals
│   ├── toolDefinitionMap.ts    # Auto-generated MCP tool definitions (don’t edit by hand)
│   ├── toolUtils.ts            # Helpers to build MCP tool objects (metadata, descriptions)
│   ├── types.ts                # Core TS interfaces (MCP, environment config, etc.)
│   │
│   ├── server/                 # MCP server logic
│   │   ├── index.ts            # Create & start the MCP server (Server + Stdio transport)
│   │   ├── handlers.ts         # `handleListTools` & `handleCallTool` implementations
│   │   ├── visualization.ts    # Post-process & save image/video outputs (masks, boxes, depth maps)
│   │   └── config.ts           # Load & validate .env, export SERVER_CONFIG & EnvConfig
│   │
│   ├── utils/                  # Generic utilities
│   │   ├── file.ts             # File handling (base64 encode images/PDFs, read streams)
│   │   └── http.ts             # Axios wrappers & error formatting
│   │
│   └── validation/             # Zod schema generation & argument validation
│       └── schema.ts           # Convert JSON Schema → Zod, validate incoming tool args
│
├── build/                      # Compiled JavaScript (generated after `npm run build`)
│   ├── index.js
│   ├── generateTools.js
│   ├── toolDefinitionMap.js
│   └── …                       # Mirror of `src/` structure
│
├── output/                     # Runtime artifacts (bounding boxes, masks, depth maps, etc.)
│
└── assets/                     # Static assets (e.g., demo.gif)
    └── demo.gif

🔍 Key Components

1. `src/generateTools.ts`

2. `src/toolDefinitionMap.ts`

3. `src/server/handlers.ts`

4. `src/server/visualization.ts`

5. `src/utils/file.ts`

6. `src/utils/http.ts`

7. `src/validation/schema.ts`

8. `src/index.ts`

vision-tools-api MCP Server (v0.1.0) running on stdio, proxying to https://api.va.landing.ai

🚧 Error Handling & Logs

• Validation Errors

If you send invalid or missing parameters, the server returns:

{
    "id": 3,
    "error": {
      "code": -32602,
      "message": "Validation error: missing required parameter ‘imagePath’"
    }
  }

• Network Errors

Axios errors (timeouts, 5xx) are caught and returned as:

{
    "id": 4,
    "error": {
      "code": -32000,
      "message": "VisionAgent API error: 502 Bad Gateway"
    }
  }

• Internal Exceptions

Uncaught exceptions in handlers produce:

{
    "id": 5,
    "error": {
      "code": -32603,
      "message": "Internal error: Unexpected token in JSON at position 345"
    }
  }

🛟 Troubleshooting

Authentication failed

• Verify `VISION_AGENT_API_KEY` is correct and active.
• Free tiers have rate limits—check your dashboard.
• Ensure outbound HTTPS to `api.va.landing.ai` isn’t blocked by a proxy/VPN.

“Tool not found” in chat

The local tool map may be stale. Run:

npm run generate-tools
npm start

Node < 20 error

The code uses the Blob & FormData APIs natively introduced in Node 20.

Upgrade via `nvm install 20` (mac/Linux) or download from nodejs.org if on Windows.

For other issues, refer to the MCP documentation: https://modelcontextprotocol.io/quickstart/user

Also not that specific clients will have their own helpful documentation. For example, if you are using the OpenAI Agents SDK, refer to their documentation here: https://openai.github.io/openai-agents-python/mcp/

🤝 Contributing

We love PRs!

1. Fork → `git checkout -b feature/my-feature`.

2. `npm run typecheck` (no errors)

3. Open a PR explaining what and why.

🔒 Security & Privacy

• The MCP server runs locally, so no files are forwarded anywhere except Landing AI’s API endpoints you explicitly call.
• Output images/masks are written to `OUTPUT_DIRECTORY` only on your machine.
• No telemetry is collected by this project.

> *Made with ❤️ by the LandingAI Team.*