PicDefense.io MCP Server
# PicDefense.io MCP Server
A [Model Context Protocol](https://modelcontextprotocol.io) server for the
[PicDefense.io](https://picdefense.io) API β let AI agents run reverse-image risk
analysis, EXIF extraction, image backlink discovery, and image content
detection (face / landmark / logo / label / SafeSearch) on any image URL.
## Features
- π **Per-user authentication** β each connection carries its own PicDefense API token; the server holds no keys
- π **Dual transport** β modern Streamable HTTP (`/mcp`) and legacy SSE (`/sse`)
- π§° **11 tools** covering the full PicDefense API v2
- π³ **Docker-ready** β production container behind nginx
- π **Built-in docs** β Swagger UI at `/docs`
## Tools
| Tool | Description |
|------|-------------|
| `picdefense_get_credits` | Remaining account credit balance |
| `picdefense_check_image_risk` | Reverse-image risk analysis + **picrisk** score (core tool) |
| `picdefense_extract_exif` | Extract EXIF metadata (camera, timestamps, GPS) |
| `picdefense_detect_face` | Detect a human face in an image |
| `picdefense_detect_landmark` | Detect a recognizable landmark |
| `picdefense_detect_logo` | Detect a brand logo |
| `picdefense_safesearch` | Content-safety (adult/violence/racy/β¦) assessment |
| `picdefense_find_backlinks` | Find pages where an image appears |
| `picdefense_detect_labels` | Detect descriptive labels for image contents |
| `picdefense_extract_text` | Extract text from an image via OCR |
| `picdefense_detect_watermark` | Detect a visible stock/photographer watermark (source + confidence) |
All image tools take a single `url` (a public http/https image URL). **Most tools
consume account credits per call** β use `picdefense_get_credits` to check your balance.
## Authentication
Every request authenticates with your PicDefense **API token**, which is your
**user id and API key joined by a colon**:
```
USERID:APIKEY
```
Find both in your PicDefense.io account settings:
**https://app.picdefense.io/?returnUrl=https://app.picdefense.io/dashboard/settings**
The token is sent as the `X-API-TOKEN` header to the API (`https://app.picdefense.io/api/v2`).
## Quick start
### Hosted server (recommended)
The hosted server runs at **https://mcp.picdefense.io**. Add it to Claude Code:
```bash
# Streamable HTTP (recommended)
claude mcp add -t http picdefense "https://mcp.picdefense.io/mcp" \
--header "X-API-Token: USERID:APIKEY"
# or SSE
claude mcp add -t sse picdefense "https://mcp.picdefense.io/sse?token=USERID:APIKEY"
```
Quick HTTP smoke test:
```bash
curl -X POST https://mcp.picdefense.io/mcp \
-H "Content-Type: application/json" \
-H "X-API-Token: USERID:APIKEY" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'
```
### Claude Desktop (hosted)
Claude Desktop launches MCP servers as local commands, so reach the hosted server
through the [`mcp-remote`](https://www.npmjs.com/package/mcp-remote) bridge (requires
Node.js installed). See [`claude_desktop_config.example.json`](claude_desktop_config.example.json):
```json
{
"mcpServers": {
"picdefense": {
"command": "npx",
"args": [
"-y",
"mcp-remote",
"https://mcp.picdefense.io/sse?token=YOUR_USERID:YOUR_APIKEY"
]
}
}
}
```
Edit your `claude_desktop_config.json` (Settings β Developer β Edit Config), add the
`mcpServers` block above with your `USERID:APIKEY`, then fully quit and reopen Claude Desktop.
> Testing against a plain-HTTP server (e.g. `http://<host>:6910`) instead of HTTPS?
> `mcp-remote` blocks non-HTTPS origins unless the host is `localhost` β append
> `"--allow-http"` to the `args` array, or reach it over an SSH tunnel to `localhost`.
### Local (stdio) via npx β no clone needed
Run the published package directly. Requires Node.js installed.
```json
{
"mcpServers": {
"picdefense": {
"command": "npx",
"args": [
"-y",
"@picdefenseio/mcp-server",
"--api-token",
"USERID:APIKEY"
]
}
}
}
```
You can also pass the token via the `PICDEFENSE_API_TOKEN` env var instead of `--api-token`.
### Local (stdio) from source
Clone and build, then point Claude Desktop at the built entry point:
```bash
git clone https://github.com/rchanllc/picdefenseio-mcp-server.git
cd picdefenseio-mcp-server
npm install
npm run build
```
Then use `"command": "node"` with `"args": ["/absolute/path/to/dist/index.js", "--api-token", "USERID:APIKEY"]`.
## Configuration
| Variable | Default | Description |
|----------|---------|-------------|
| `PICDEFENSE_API_TOKEN` | β | `USERID:APIKEY` (stdio only; hosted server reads it per-connection) |
| `PICDEFENSE_API_BASE_URL` | `https://app.picdefense.io/api/v2` | API base URL |
| `PORT` | `6910` | Hosted server listen port |
## Running the hosted server
### Development
```bash
npm run dev:sse # tsx watch, auto-reload
```
### Production (Docker)
```bash
# via docker compose
docker compose up --build -d
# or the helper script (handles build + health check)
./deploy.sh
```
The container is named `picdefenseio_mcp` and listens on port **6910**.
### HTTP endpoints
| Method | Path | Purpose |
|--------|------|---------|
| GET | `/health` | Health check (used by Docker + CI) |
| GET | `/api/info` | Server + transport info |
| GET | `/tools` | List available tools |
| GET | `/docs` | Swagger UI for the underlying API |
| GET | `/sse?token=USERID:APIKEY` | Open an SSE MCP session |
| POST | `/messages?sessionId=<id>` | SSE session message channel |
| POST | `/mcp` | Streamable HTTP MCP (header `X-API-Token`) |
## Self-hosting
There is no CI/CD in this repo β host it yourself. On your server:
```bash
git clone https://github.com/rchanllc/picdefenseio-mcp-server.git
cd picdefenseio-mcp-server
docker compose up --build -d # or: ./deploy.sh
```
The container is named `picdefenseio_mcp` and listens on **6910**. To update,
`git pull` and re-run `docker compose up --build -d`.
Front it with nginx at `https://mcp.picdefense.io` β `127.0.0.1:6910`
(`proxy_buffering off` and a long read timeout are recommended for the `/sse` path).
## Architecture
```
ββββββββββββββββ ββββββββββββββββββββββββββ βββββββββββββββββββββββββββ
β MCP Client βββββΆβ PicDefense MCP Server βββββΆβ PicDefense.io API v2 β
β (Claude etc.)β β (port 6910) β β app.picdefense.io/api/v2β
ββββββββββββββββ ββββββββββββββββββββββββββ βββββββββββββββββββββββββββ
token (USERID:APIKEY) forwarded as X-API-TOKEN ββββββββββββΆ
```
Each connection builds its own API client + MCP server bound to the caller's token,
so the service is multi-tenant and stateless with respect to credentials.
## License
MIT β see [LICENSE](LICENSE).
TDQS
Scored across 11 tools
Each tool targets a distinct aspect of image analysis: risk, EXIF, faces, landmarks, logos, content safety, backlinks, labels, OCR, watermark, and credits. The descriptions clearly delineate overlapping concepts like reverse-image-search findings vs. dedicated detection tools, so an agent can unambiguously select the right one.
All tool names follow the 'picdefense_<verb_noun>' pattern (e.g., extract_exif, detect_face, find_backlinks) with a consistent prefix, but 'safesearch' deviates as a single compound noun rather than a verb_noun construction. This minor inconsistency does not hinder readability.
With 11 tools, the server achieves a well-scoped balanceβenough to cover the full spectrum of image risk and detection features without overwhelming the agent. Each tool earns its place and there is no apparent bloat.
The tool surface comprehensively covers the image-risk domain: credit management, core risk scoring, EXIF, face/landmark/logo/watermark detection, content safety, backlinks, labels, and OCR. No obvious lifecycle or analytical gaps exist for typical use cases.