media-mcp (Node.js)

Deployment update (2026-09-06): the SAM3 source and release scripts are now maintained in sam3-http-server, rooted at /opt/ai/sam3. API schema candidate head is 20260906_0007. SAM3 production bootstrap and nine-tool E2E remain pending. See the four-project rollout.

English | 中文

A video enhancement, image enhancement/colorization/denoising, and image segmentation service based on the MCP protocol, acting as an MCP Client-Server to interact with backend HTTP Servers.

Release status (2026-09-06): npm latest is 0.2.1; that published artifact contains 5 tools (3 video + 2 SAM3) and its runtime metadata incorrectly reports 0.3.0. This repository is the unreleased 0.3.0 candidate, which fixes version consistency and adds 4 image tools plus an optional IMAGE_API_BASE_URL override. Image and video share the same public HTTP server and /enhance base URL; the candidate backend keeps their prepare concurrency, Redis queues, workers, and processing concurrency separate. The image routes now exist in candidate code but are not deployed or real-AI/TOS verified, so the image tools below are not a production npm capability until all 9 tools pass smoke tests and 0.3.0 is published.

Features

Provides the following MCP Tools:

Video Enhancement

• create_task - Create a video enhancement task (supports URL or local file upload)
• get_task_status - Query task status
• enhance_video_sync - Synchronously enhance video (blocking wait, truncated at ~45s by default)

Image Enhancement (unreleased 0.3.0 candidate)

• enhance_image_sync - Enhance image quality and optimize faces (supports URL or local file upload)
• colorize_image_sync - Colorize black-and-white photos (supports URL or local file upload)
• denoise_image_sync - Remove noise from images (supports URL or local file upload)
• get_image_task_status - Query image task status (for polling after sync timeout)

Image Segmentation (SAM3)

• sam3_predict - SAM3 image segmentation (supports local path, URL, or Base64 image)
• get_sam3_task_status - Query SAM3 task status (for polling after sync timeout)

Prerequisites

• Node.js >= 18 (check: node --version)
• API Key (required for authentication)

Lazy Install (Recommended)

If your AI Agent has a known MCP config path, just copy the line below and send it to your AI:

Install the npm package @avclabs.ai/media-mcp as an MCP server. My API Key is: sk-xxxxxxxx.

The AI will automatically:

1. Detect your MCP client
2. Find the config file path
3. Write the correct configuration
4. Prompt you to restart the client

Manual Install

No installation needed. Use npx directly in your MCP client config.

1. Claude Code (CLI)

Run in Claude Code:

/mcp

Check the output for the "User MCPs" section to find the config file path, then edit that file.

Common paths (if /mcp is unavailable):

• Windows: %USERPROFILE%\.claude.json
• macOS: ~/.claude.json
• Linux: ~/.claude.json
• Legacy/Alternative: ~/.claude/mcp.json

Paste this (replace your-api-key):

{
  "mcpServers": {
    "video-enhancement": {
      "command": "npx",
      "args": ["-y", "@avclabs.ai/media-mcp@latest"],
      "env": {
        "API_KEY": "your-api-key"
      }
    }
  }
}

Save and run /mcp to verify it's loaded.

2. Cursor

Go to Settings > Tools & MCPs > Add New MCP Server:

• Name: video-enhancement
• Type: command
• Command:
  bash

  env API_KEY=your-api-key npx -y @avclabs.ai/media-mcp@latest
  

Or edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "video-enhancement": {
      "command": "npx",
      "args": ["-y", "@avclabs.ai/media-mcp@latest"],
      "env": {
        "API_KEY": "your-api-key"
      }
    }
  }
}

Verify Installation

After restarting your client, check if the tools are available:

1. Or ask: "What tools do you have available?"
2. With the current npm latest=0.2.1, you should see create_task, get_task_status, enhance_video_sync, sam3_predict, and get_sam3_task_status.
3. After 0.3.0 is published, you should additionally see enhance_image_sync, colorize_image_sync, denoise_image_sync, and get_image_task_status.

Configuration Options

IMAGE_API_BASE_URL is implemented by the unreleased 0.3.0 candidate as an optional override. Production uses the shared /enhance service, so it should normally be omitted; the candidate then resolves image and video calls to the same base URL.

Custom Endpoint (0.3.0 candidate)

{
  "env": {
    "HTTP_API_BASE_URL": "https://your-media-endpoint.com",
    "API_KEY": "your-api-key",
    "SAM3_API_BASE_URL": "https://your-sam3-endpoint.com"
  }
}

Or via CLI args:

npx -y @avclabs.ai/media-mcp@0.3.0 --base-url https://your-media-endpoint.com --api-key your-api-key --sam3-base-url https://your-sam3-endpoint.com

Run this command only after 0.3.0 is published. The optional --image-base-url remains available only for deployments that intentionally split the public image endpoint; npm 0.2.1 does not support it.

Recommended Workflow

This project provides both synchronous and asynchronous modes.

Because MCP Agents typically enforce a ~60-second timeout per tool call, tasks with longer processing times (video enhancement) are strongly recommended to use asynchronous mode:

Asynchronous Mode (Recommended)

Video Enhancement:

1. Call create_task to create a task → immediately get task_id
2. Wait a few seconds, then call get_task_status to query the status
3. If status is processing, continue waiting and repeat step 2
4. If status is completed, the task is done and the result contains video_url
5. If status is failed, the task failed and the result contains error_message

Synchronous Mode (Simple Scenarios)

Video Enhancement:

• Call enhance_video_sync → the server polls internally
• Defaults to a maximum wait of 45 seconds
• If completed within 45 seconds, returns the result directly
• If not completed within 45 seconds, returns task_id and instructions for the Agent to switch to get_task_status

Image Segmentation (SAM3):

• Call sam3_predict → the server polls internally
• Defaults to at most 25 status queries within a ~50-second budget (25 × 2-second interval; whichever limit is reached first ends the wait)
• If the task completes within the budget, returns the segmentation result directly
• If not completed within the budget, returns a truncation notice (with task_id) indicating the task is still processing

Usage Examples

Once configured, ask your AI agent naturally:

"Enhance this video to 1080p: https://example.com/video.mp4"

"Improve the quality of /Users/me/Desktop/video.mp4 to 2k"

"Enhance this image: https://example.com/photo.jpg"

"Colorize this black-and-white photo: /Users/me/Desktop/old_photo.png"

"Remove noise from this image: C:\Users\xxx\noisy.jpg"

"Analyze this image and find all objects: C:\Users\xxx\photo.png"

"Use SAM3 to segment this image, prompt: 'find all cars'"

The agent will automatically choose sync or async tools based on task complexity.

Provided Tools

Video Enhancement

create_task

Create an asynchronous video enhancement task.

Recommended for most use cases. Ideal for longer videos (over 10 seconds) to avoid timeouts and blocking the connection.

Returns:

{
  "success": true,
  "task_id": "xxx",
  "status": "processing"
}

get_task_status

Query video enhancement task status.

The returned status field can be: processing, completed, or failed. If status is processing, you need to wait a few seconds and call this tool again.

Returns:

{
  "success": true,
  "task_id": "xxx",
  "status": "completed",
  "progress": 100,
  "video_url": "https://...",
  "message": "Task is still processing, please check again later"
}

The message field only appears when status is processing, prompting the Agent to continue waiting.

enhance_video_sync

Synchronously enhance video (blocks until completion).

Best for short videos (estimated processing time < 1 minute). If the task is not completed within the default 45-second wait budget, the tool returns early with a task_id, and you need to use get_task_status to continue querying.

Truncated return example (not completed within the default 45-second budget):

{
  "success": true,
  "status": "processing",
  "task_id": "xxx",
  "message": "Task is still processing (waited 45 seconds, 9 status queries). Please use get_task_status to continue polling.",
  "note": "The synchronous wait for this long-running task has been truncated. Switch to get_task_status polling."
}

Image Enhancement

Three image processing tools are provided, each targeting a specific use case:

All three tools share the same parameters and behavior pattern. They are synchronous — the tool blocks until the image is processed or the timeout is reached.

Supported image formats: PNG, JPG, JPEG, BMP, WebP, etc.

Two upload methods:

1. URL upload: provide a publicly accessible image URL (type: "url")
2. Local upload: provide a local file path, the MCP Server auto-uploads to TOS object storage (type: "local", max file size: 100MB)

enhance_image_sync

Synchronously enhance an image to improve quality and optimize faces.

The tool internally creates a task and polls for the result. If processing completes within the timeout (default 45s), the result is returned directly. If not, the tool returns early with a task_id — use get_image_task_status to continue polling.

Normal completion return:

{
  "success": true,
  "task_id": "xxx",
  "status": "completed",
  "progress": 100,
  "image_url": "https://..."
}

Truncated return (not completed within the default 45-second budget):

{
  "success": true,
  "status": "processing",
  "task_id": "xxx",
  "message": "Task is still processing (waited 45 seconds, 9 status queries). Please use get_image_task_status to continue polling.",
  "note": "The synchronous wait for this long-running task has been truncated. Switch to get_image_task_status polling."
}

colorize_image_sync

Synchronously colorize a black-and-white photo with AI.

Best for old black-and-white photos. The AI will add realistic colors to the image. Supports the same parameters and return format as enhance_image_sync.

Returns: Same format as enhance_image_sync.

denoise_image_sync

Synchronously remove noise from an image.

Best for grainy/noisy photos taken in low-light conditions or with high ISO settings. Supports the same parameters and return format as enhance_image_sync.

Returns: Same format as enhance_image_sync.

get_image_task_status

Query image processing task status. Used to poll for results when a sync tool times out.

The returned status field can be: processing, completed, or failed. If status is processing, wait a few seconds and call this tool again.

Returns:

{
  "success": true,
  "task_id": "xxx",
  "status": "completed",
  "progress": 100,
  "image_url": "https://...",
  "message": "Task is still processing, please check again later"
}

The message field only appears when status is processing, prompting the Agent to continue waiting.

Recommended Workflow for Image Tools

1. For most images: Call enhance_image_sync / colorize_image_sync / denoise_image_sync directly — the tool handles everything and returns the result
2. If truncated: The tool returns a task_id, then use get_image_task_status to poll until status becomes completed or failed
3. If failed: Check the error_message field for details

Image Segmentation (SAM3)

sam3_predict

Analyze an image using the SAM3 segmentation API to generate inference results (masks, boxes, scores).

Parameters:

Image input (choose one, must provide exactly one):

• imagePath (string): Absolute path of a local image file. Supports common formats (PNG, JPG, JPEG).

  • Example: "C:\\Users\\xxx\\photo.png", "/home/user/images/cat.jpg"
  • Use when: The user explicitly provides a local file path

• imageUrl (string): Publicly accessible URL of the image.

  • Example: "https://example.com/photo.jpg"
  • Use when: The image is already online and the user provides a link
  • Note: The URL must be publicly accessible. Links requiring login or signatures are not supported

• imageBase64 (string): Base64-encoded image data.

  • Example: "iVBORw0KGgoAAAANSUhEUgAA..."
  • Use when: The user drags or uploads an image attachment, and the Agent encodes it as base64
  • Note: Large images will produce very large base64 strings, which may slow transmission

Other parameters:

• prompt (string, required): English text prompt specifying the target object to segment. Since the SAM3 model only accepts English prompts, provide an English description. If the user provides Chinese or other non-English text, the Agent will automatically translate it before calling the tool.

Normal completion return:

After inference completes, returns a JSON string containing three fields:

• masks: 2D array. Each element is a binary mask (values 0 or 1) with the same dimensions as the input image, marking the pixel-level location of detected objects. The i-th mask corresponds to the i-th detected object instance.

• boxes: 2D array. Each element is a bounding box in [x1, y1, x2, y2] format, representing the rectangular region of the detected object. x1, y1 are the top-left coordinates; x2, y2 are the bottom-right coordinates.

  Coordinate system: The top-left corner of the image is the origin (0, 0). The x-axis increases to the right, and the y-axis increases downward, in pixels. For example, [120, 80, 300, 450] means the region starts 120px from the left edge and 80px from the top edge, extending to 300px from the left and 450px from the top. Width = x2 - x1 = 180px, Height = y2 - y1 = 370px.

• scores: 1D array. Each element is a confidence score for the corresponding detection result, ranging from 0 to 1. Higher scores indicate greater model confidence.

Example result JSON:

{
  "masks": [
    [[0, 0, 1, ...], [0, 1, 1, ...], ...],
    [[0, 0, 0, ...], [0, 0, 1, ...], ...]
  ],
  "boxes": [
    [120, 80, 300, 450],
    [400, 200, 600, 500]
  ],
  "scores": [0.95, 0.87]
}

Truncated return example (not completed within 50s):

{
  "success": true,
  "status": "processing",
  "task_id": "xxx",
  "message": "Task is still processing (waited 50 seconds, 25 status queries). Please use get_sam3_task_status to continue polling.",
  "note": "The synchronous wait for this long-running task has been truncated. Switch to task status polling."
}

Result download failure (completed task, missing result): If the task completed but the result file could not be downloaded, the tool returns success: false with the task_id and status: "completed". Use get_sam3_task_status with that task_id to recover the result. Do not resubmit the task.

{
  "success": false,
  "task_id": "xxx",
  "status": "completed",
  "error": "Failed to download the completed result for task xxx: ...",
  "note": "The task has completed; only the result download failed. Do not resubmit. Use get_sam3_task_status with this task_id to recover the result."
}

get_sam3_task_status

Query SAM3 segmentation task status. Used to poll for results when sam3_predict times out.

The returned status field can be: processing, completed, or failed. If status is processing, wait a few seconds and call this tool again.

Completed return:

{
  "success": true,
  "task_id": "xxx",
  "status": "completed",
  "result_url": "https://..."
}

Processing return:

{
  "success": true,
  "task_id": "xxx",
  "status": "processing",
  "message": "Task is still processing, please check again later."
}

Failed return:

{
  "success": false,
  "task_id": "xxx",
  "status": "failed",
  "error": "Task failed"
}

FAQ

Agent reports timeout when calling tools?

This is the primary issue this project addresses. MCP Agents (such as Claude, Cursor) typically enforce a ~60-second timeout per tool call. If task processing exceeds this limit, the Agent will error and disconnect.

Solutions:

1. Prefer asynchronous tools: For video enhancement and other time-consuming tasks, always use create_task + get_task_status. These tools return instantly on each call and will not trigger timeouts.

2. Sync tool truncation mechanism: enhance_video_sync has an internal 45-second truncation limit (default timeout). If the task is not completed within the budget, the tool proactively returns a task_id and instructs the Agent to use get_task_status to follow up.

3. SAM3 truncation mechanism: sam3_predict is bounded by two independent limits derived from its config: at most SAM3_POLL_MAX_ATTEMPTS status queries (default 25) and a time budget of SAM3_POLL_INTERVAL_MS × attempts (default 25 × 2 s ≈ 50 s). Whichever runs out first truncates the wait; the returned notice carries the task_id and the actual elapsed time and query count, and the task keeps running remotely.

4. Adjust SAM3 polling parameters (advanced): The two SAM3 limits are independent, so you can trade interval for attempt count freely (e.g. SAM3_POLL_INTERVAL_MS=1000 with SAM3_POLL_MAX_ATTEMPTS=50 still gives a ~50-second budget with twice as many queries). Lower the interval if the task usually finishes in a few seconds and you want quicker completion detection. Invalid explicit values (non-integer or outside the allowed range) abort startup before any task can be submitted. The wait budget covers only status polling after the task is created — source preparation, upload, task submission, and the final result download are not part of it, so the total tool call can take longer than the budget.

Drag-and-drop attachment says file not found?

This is a known limitation of stdio MCP. When dragging or uploading an attachment through the Agent interface, the file path is usually not automatically passed to the MCP Server.

Solutions:

1. Provide the path simultaneously (recommended): After dragging the image, add the local absolute path in your message:

   "Please analyze this image D:\\photos\\cat.jpg and find the cat"

2. Wait for auto-encoding: Claude may automatically encode the image as base64. If successful, no extra action is needed.

3. Reply to path inquiry: If Claude asks for the image path, simply reply with the local absolute path.

Is there a priority among the three input methods?

There is no strict priority. Claude will automatically choose the most appropriate method based on conversation context:

• You provided a local path → uses imagePath
• You provided a web link → uses imageUrl
• You dragged an attachment without a path → tries imageBase64

What image formats are supported?

Common formats: PNG, JPG, JPEG, BMP, WebP, etc. PNG or JPG is recommended.

What if URL image download fails?

Ensure the URL is publicly accessible, requiring no login, cookies, or signatures. If the image is on a service requiring authentication (e.g., private S3 Bucket, login-required image host), download it locally first and use imagePath.

What if the base64 image is too large?

If the image is very large (e.g., 4K resolution), the base64-encoded data will be very large and may slow transmission. Suggestions:

1. Use imagePath instead
2. Or compress the image before encoding

File Upload Notes

When type is "local":

1. File is read locally by the MCP Server
2. Uploaded directly to TOS object storage via pre-signed URL
3. Max file size: 100MB

Troubleshooting

"command not found: npx"

Install Node.js >= 18: https://nodejs.org/

"API key is required: pass --api-key or set the API_KEY environment variable"

Your API Key is missing. Double-check the env.API_KEY in your config.

MCP Server shows red/error in client

Check logs:

• Claude Desktop macOS: ~/Library/Logs/Claude/mcp*.log
• Claude Desktop Windows: %APPDATA%\Claude\logs\mcp*.log
• Cursor: Output panel > MCP

"TOS upload failed"

Usually a signature mismatch. Ensure your IMAGE_API_BASE_URL (or its HTTP_API_BASE_URL fallback) and API_KEY are correct and active.

Global Install (Alternative)

If you prefer not using npx every time:

npm install -g @avclabs.ai/media-mcp

Then use "command": "media-mcp" with "args": ["--api-key", "your-api-key"] in your config.

Development and Release

This package runs locally in the MCP client and is released through npm; it is not a remote Node daemon. The sibling media-mcp-api-http-server is live as the shared /enhance video/image/account owner, although its current production release only implements video routes; candidate code adds the image routes, durable prepare/dispatch outboxes, task-level credit idempotency, and separate image/video queues. The Portal is live and SAM3 remains an external production dependency. Real image AI/TOS E2E, production backend rollout, and SAM3 JSON health are still pending, so 0.3.0 must not be published yet. Before publishing, run:

npm ci
npm run release:verify

See the release guide for version synchronization, publish order, smoke tests, and the relationship with the portal/backend deployment.

License

MIT License - See LICENSE file for details