What you get
Once configured, Claude Code gains nine native tools it calls automatically when a task involves video or images:
| Tool | What it does |
|---|---|
create_task | Submits an async video enhancement task (upscaling, quality enhancement) |
get_task_status | Polls task progress and returns the result URL when done |
enhance_video_sync | Synchronous enhancement for short clips (45s truncation limit) |
enhance_image_sync | Synchronous image enhancement and 1–4x upscaling |
colorize_image_sync | Colorizes black-and-white photos |
denoise_image_sync | Removes image noise and grain |
get_image_task_status | Polls image task progress and returns the result URL |
sam3_predict | SAM3 segmentation: text prompt → masks, boxes, confidence scores |
get_sam3_task_status | Polls SAM3 task progress and returns the result URL |
Prefer a smaller footprint? The server is being split into per-capability packages (video, image, SAM3) — once published, installing only the capability you need keeps fewer tool definitions in the session context and reduces tool-selection ambiguity.
Setup in one command
Paste this into a Claude Code session and confirm:
Install @avclabs.ai/media-mcp as an MCP server
Or add it manually to ~/.claude.json (Windows: %USERPROFILE%\.claude.json):
{
"mcpServers": {
"media-mcp": {
"command": "npx",
"args": [
"-y",
"-p",
"@avclabs.ai/media-mcp@latest",
"media-mcp"
],
"env": {
"API_KEY": "your-api-key"
}
}
}
}Requires Node.js ≥ 18 and an API key from the dashboard (free quota included, no credit card).
Example prompts
- "Upscale this video to 1080p: https://example.com/demo.mp4"
- "Enhance the quality of screen-recording.mp4 and notify me when it's done"
- "Colorize ./assets/grandparents-1960.jpg and upscale it 2x"
- "Find and mask all people in ./assets/group-photo.jpg"
Claude picks async mode for long videos automatically, polls the task, and hands you the result URL when processing finishes.
Why MCP instead of raw API calls
With a raw HTTP API you write the polling loop, error handling and file handling yourself. With the MCP server, Claude Code owns the workflow: it chooses sync vs async mode, retries with get_task_status when a call truncates, and chains segmentation results into the next step of your task — all from a single sentence.