DINO-X Image Detection MCP Server
README.md
# DINO-X MCP Server
[](https://opensource.org/licenses/Apache-2.0) [](https://www.npmjs.com/package/@deepdataspace/dinox-mcp) [](https://www.npmjs.com/package/@deepdataspace/dinox-mcp) [](https://github.com/IDEA-Research/DINO-X-MCP/pulls) [](https://lobehub.com/mcp/idea-research-dino-x-mcp) [](https://github.com/IDEA-Research/DINO-X-MCP/stargazers)
**English** | [δΈζ](README_ZH.md)
DINO-X Official MCP Server β powered by the DINO-X and Grounding DINO models β brings fine-grained object detection and image understanding to your multimodal applications.
<p align="center">
<video width="800" controls>
<source src="https://dds-frontend.oss-cn-shenzhen.aliyuncs.com/dinox-mcp/dinox-mcp-en-overveiw.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</p>
## Why DINO-X MCP?
With DINO-X MCP, you can:
- Fine-Grained Understanding: Full image detection, object detection, and region-level descriptions.
- Structured Outputs: Get object categories, counts, locations, and attributes for VQA and multi-step reasoning tasks.
- Composable: Works seamlessly with other MCP servers to build end-to-end visual agents or automation pipelines.
## Transport Modes
DINO-X MCP supports two transport modes:
| Feature | STDIO (default) | Streamable HTTP |
|:--|:--|:--|
| Runtime | Local | Local or Cloud |
| Transport | Standard I/O | HTTP (streaming responses) |
| Input source | `file://` and `https://` | `https://` only |
| Visualization | Supported (saves annotated images locally) | Not supported (for now) |
## Quick Start
### 1. Prepare an MCP client
Any MCP-compatible client works, e.g.:
- [Cursor](https://www.cursor.com/)
- [WindSurf](https://windsurf.com/)
- [Trae](https://www.trae.ai/)
- [Cherry Studio](https://www.cherry-ai.com/)
### 2. Get your API key
Apply on the DINO-X platform: [Request API Key](https://cloud.deepdataspace.com/request_api) (new users get free quota).
### 3. Configure MCP
#### Option A: Official Hosted Streamable HTTP (Recommended)
Add to your MCP client config and replace with your API key:
```json
{
"mcpServers": {
"dinox-mcp": {
"url": "https://mcp.deepdataspace.com/mcp?key=your-api-key"
}
}
}
```
#### Option B: Use the NPM package locally (STDIO)
Install Node.js first
- Download the installer from [nodejs.org](https://nodejs.org/)
- Or use command:
```bash
# macOS / Linux
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash
# or
wget -qO- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash
# load nvm into current shell (choose the one you use)
source ~/.bashrc || true
source ~/.zshrc || true
# install and use LTS Node.js
nvm install --lts
nvm use --lts
# Windows (one of the following)
winget install OpenJS.NodeJS.LTS
# or with Chocolatey (in admin PowerShell)
iwr -useb https://raw.githubusercontent.com/chocolatey/chocolatey/master/chocolateyInstall/InstallChocolatey.ps1 | iex
choco install nodejs-lts -y
```
Configure your MCP client:
```json
{
"mcpServers": {
"dinox-mcp": {
"command": "npx",
"args": ["-y", "@deepdataspace/dinox-mcp"],
"env": {
"DINOX_API_KEY": "your-api-key-here",
"IMAGE_STORAGE_DIRECTORY": "/path/to/your/image/directory"
}
}
}
}
```
Note: Replace `your-api-key-here` with your real key.
#### Option C: Run from source locally
Make sure Node.js is installed (see Option B), then:
```bash
# clone
git clone https://github.com/IDEA-Research/DINO-X-MCP.git
cd DINO-X-MCP
# install deps
npm install
# build
npm run build
```
Configure your MCP client:
```json
{
"mcpServers": {
"dinox-mcp": {
"command": "node",
"args": ["/path/to/DINO-X-MCP/build/index.js"],
"env": {
"DINOX_API_KEY": "your-api-key-here",
"IMAGE_STORAGE_DIRECTORY": "/path/to/your/image/directory"
}
}
}
}
```
## CLI Flags & Environment Variables
- Common flags
- `--http`: start in Streamable HTTP mode (otherwise STDIO by default)
- `--stdio`: force STDIO mode
- `--dinox-api-key=...`: set API key
- `--enable-client-key`: allow API key via URL `?key=` (Streamable HTTP only)
- `--port=8080`: HTTP port (default 3020)
- Environment variables
- `DINOX_API_KEY` (required/conditionally required): DINO-X platform API key
- `IMAGE_STORAGE_DIRECTORY` (optional, STDIO): directory to save annotated images
- `AUTH_TOKEN` (optional, HTTP): if set, client must send `Authorization: Bearer <token>`
Examples:
```bash
# STDIO (local)
node build/index.js --dinox-api-key=your-api-key
# Streamable HTTP (server provides a shared API key)
node build/index.js --http --dinox-api-key=your-api-key
# Streamable HTTP (custom port)
node build/index.js --http --dinox-api-key=your-api-key --port=8080
# Streamable HTTP (require client-provided API key via URL)
node build/index.js --http --enable-client-key
```
Client config when using `?key=`:
```json
{
"mcpServers": {
"dinox-mcp": {
"url": "http://localhost:3020/mcp?key=your-api-key"
}
}
}
```
Using `AUTH_TOKEN` with a gateway that injects `Authorization: Bearer <token>`:
```bash
AUTH_TOKEN=my-token node build/index.js --http --enable-client-key
```
Client example with `supergateway`:
```json
{
"mcpServers": {
"dinox-mcp": {
"command": "npx",
"args": [
"-y",
"supergateway",
"--streamableHttp",
"http://localhost:3020/mcp?key=your-api-key",
"--oauth2Bearer",
"my-token"
]
}
}
}
```
## Tools
| Capability | Tool ID | Transport | Input | Output |
|:--|:--|:--|:--|:--|
| Full-scene object detection | `detect-all-objects` | STDIO / HTTP | Image URL | Category + bbox + (optional) captions |
| Text-prompted object detection | `detect-objects-by-text` | STDIO / HTTP | Image URL + English nouns (dot-separated for multiple, e.g., `person.car`) | Target object bbox + (optional) captions |
| Human pose estimation | `detect-human-pose-keypoints` | STDIO / HTTP | Image URL | 17 keypoints + bbox + (optional) captions |
| Visualization | `visualize-detection-result` | STDIO only | Image URL + detection results array | Local path to annotated image |
## π¬ Use Cases
| π― Scenario | π Input | β¨ Output |
|---------|---------|---------|
| **Detection & Localization** | **π¬ Prompt:**<br>`Detect and visualize the `<br>`fire areas in the forest `<br><br>**πΌοΈ Input Image:**<br>| |
| **Object Counting** | **π¬ Prompt:**<br>`Please analyze this`<br>`warehouse image, detect`<br>`all the cardboard boxes,`<br>`count the total number`<br><br>**πΌοΈ Input Image:**<br>| <img width="1276" alt="2-2" src="https://github.com/user-attachments/assets/3f18ef8c-5e89-45fc-bd0f-f23381304272" />|
| **Feature Detection** | **π¬ Prompt:**<br>`Find all red cars`<br>`in the image`<br><br>**πΌοΈ Input Image:**<br>||
| **Attribute Reasoning** | **π¬ Prompt:**<br>`Find the tallest person`<br>`in the image, describe`<br>`their clothing`<br><br>**πΌοΈ Input Image:**<br> |  |
| **Full Scene Detection** | **π¬ Prompt:**<br>`Find the fruit with`<br>`the highest vitamin C`<br>`content in the image`<br><br>**πΌοΈ Input Image:**<br>| <br><br>*Answer: Kiwi fruit (93mg/100g)* |
| **Pose Analysis** | **π¬ Prompt:**<br>`Please analyze what`<br>`yoga pose this is`<br><br>**πΌοΈ Input Image:**<br> ||
## FAQ
- Supported image sources?
- STDIO: `file://` and `https://`
- Streamable HTTP: `https://` only
- Supported image formats?
- jpg, jpeg, webp, png
## Development & Debugging
Use watch mode to auto-rebuild during development:
```bash
npm run watch
```
Use MCP Inspector for debugging:
```bash
npm run inspector
```
## License
Apache License 2.0
TDQS
A3.6/5.0
Scored across 4 tools
Disambiguation5/5
Each tool has a clearly distinct purpose: detect-all-objects performs general object detection, detect-human-pose-keypoints focuses on human pose analysis, detect-objects-by-text enables text-guided detection, and visualize-detection-result handles visualization. There is no overlap in functionality, making tool selection straightforward for an agent.
Naming Consistency5/5
All tool names follow a consistent verb_noun pattern with hyphens (e.g., detect-all-objects, detect-human-pose-keypoints). The naming is predictable and readable throughout, with no deviations in style or convention.
Tool Count5/5
With 4 tools, the server is well-scoped for image detection tasks. Each tool earns its place by covering distinct aspects: detection (general, pose-specific, text-guided) and visualization, avoiding bloat while providing essential functionality.
Completeness4/5
The tool set covers core detection workflows (general, pose, text-guided) and visualization, with no obvious dead ends. A minor gap exists in lacking tools for modifying or deleting detection results, but agents can work around this by re-running detections or handling data externally.
Maintenance
ActivityNo data
ResponsivenessNo issues