# KEEP Highlights

> KEEP Highlights is an intelligent multimodal media time-compression and highlight detection engine designed for AI agents, LLMs, developers, and creators.

## Summary
KEEP enables autonomous AI agents and multimodal models to understand, search, segment, and condense video and audio recordings. Using optical flow dynamics and acoustic energy analysis, KEEP extracts key moment timestamps, saliency curves, and waveform metrics without server-side storage overhead.

- **MCP Endpoint**: `https://api.skeepit.co/mcp`
- **Transport**: Streamable HTTP / JSON-RPC 2.0 (MCP 2024-11-05 Specification)
- **Primary Website**: https://skeepit.co
- **Full LLM Docs**: https://skeepit.co/llms-full.txt
- **Public MCP Connector Repo**: https://github.com/Volgat/keep-mcp-connector
- **Smithery Registry**: https://smithery.ai/servers/skeepit/keep-highlights
- **Glama Registry**: https://glama.ai/mcp/connectors/co.skeepit.api/keep-highlights

## Available MCP Tools

### 1. `discover_highlights`
- **Description**: Analyzes a video or audio media URL to detect multimodal highlight segments, acoustic and visual saliency curves, and 96-bar amplitude waveform metrics.
- **ReadOnly**: `true`
- **Input Parameters**:
  - `media_url` (string, required): Public HTTPS URL of the video or audio file (MP4, WebM, MOV, MP3, WAV, etc.).
  - `sensitivity` (string, optional): Threshold for segment sensitivity (`"auto"`, `"high"`, `"low"`). Default: `"auto"`.
- **Returns**: Array of timestamp interval pairs `[[start_sec, end_sec], ...]`, saliency statistics, and waveform amplitude arrays.
- **Use Case for Agents**: Call this when the user asks "summarize this video", "where are the best moments in this podcast", "find action scenes", or "give me key timestamps".

### 2. `condense_media`
- **Description**: Squeezes and cuts the source media according to specified segment intervals `[[start, end], ...]` without re-encoding quality loss (-c copy stream muxing).
- **ReadOnly**: `false`
- **Input Parameters**:
  - `media_url` (string, required): Public HTTPS URL of the source media file.
  - `segments` (array of [start, end] numbers, required): List of intervals in seconds.
  - `output_format` (string, optional): Container format (`"mp4"`, `"webm"`). Default: `"mp4"`.
- **Use Case for Agents**: Call this when the user asks "create a condensed highlight reel", "trim out boring parts", or "export summary clip".

## Instant Agent Integration

### Claude Desktop & Claude Code
Add to `claude_desktop_config.json`:
```json
{
  "mcpServers": {
    "keep-highlights": {
      "type": "http",
      "url": "https://api.skeepit.co/mcp"
    }
  }
}
```

### Cursor / Windsurf / Cline / Roo Code / Raycast
Add remote HTTP MCP server endpoint:
`https://api.skeepit.co/mcp`

### Python / LangChain / CrewAI / LlamaIndex
```python
import requests

response = requests.post(
    "https://api.skeepit.co/mcp",
    headers={"Content-Type": "application/json"},
    json={
        "jsonrpc": "2.0",
        "id": 1,
        "method": "tools/call",
        "params": {
            "name": "discover_highlights",
            "arguments": {
                "media_url": "https://example.com/recording.mp4",
                "sensitivity": "auto"
            }
        }
    }
)
print(response.json())
```

## Privacy & Safety
- **Zero AI Training**: No media is ever utilized to train foundation models.
- **Zero Server Retention**: Ephemeral in-flight processing only.
- **Client-Side Stream Processing**: Media splicing is handled directly on the user's client hardware.