I built an MCP server that lets AI agents investigate videos

Description and use

Urma lets AI agents retrieve captions and frames from videos, enabling evidence‑based answers when text alone is insufficient.

AI agents need to retrieve captions and visual evidence from videos to answer questions about content that is not fully captured by text alone.

Urma enables AI agents to retrieve both textual and visual evidence from videos, improving the accuracy of automated research tasks.

1. Agent calls inspect_video with a video URL or local file path. 2. Urma returns source metadata, duration, and available caption tracks. 3. Agent calls search_transcript to locate a phrase. 4. Agent calls read_transcript to get caption segments for a time interval. 5. Agent calls get_frames to retrieve JPEG frames at specific timestamps or a short burst. 6. Agent analyzes frames to confirm visual state (e.g., dark mode toggle).

Urma works through five MCP tools. An investigation begins with inspect_video, which opens a source and returns an investigation reference.

Urma records which source snapshot an investigation uses and which observations have been returned.

Urma uses source-provided captions, so videos without captions can't be searched by speech.