# VideoDB

> Build agents with visual intelligence. An applied lab for visual intelligence.

Your agent can now watch video. Ask it questions, find the exact moment, and get back a clip you can stream.

Canonical: https://videodb.io · Add to Claude: https://videodb.io/claude · Add to ChatGPT: https://videodb.io/chatgpt · Talk to us: https://videodb.io/contact.md · MCP server: https://mcp.videodb.io/mcp

1,000+ developers and agents use VideoDB every week. 300+ TB of video processed.

The page opens with an example chat in Claude or ChatGPT with VideoDB connected. Asked what a drone shot shows, the agent watches the 15-second video and answers with its 3 scenes: a drone rising over blue-green striped hills at sunrise, ridges of grey and teal clay running down from the summit, and low sun behind the flat-topped peak. Asked to find every scene with a robot in it, it searches 4 videos and returns 3 scenes. Asked to cut a 10-second clip of the robot in the market, it returns the clip, ready to share.

VideoDB is in Claude and ChatGPT. Ask it about any video: https://videodb.io/claude, https://videodb.io/chatgpt

## Teams using VideoDB every day

Capitol AI, Ezoic, Corvic AI, Virello, Clario, Muzology, BioRugged, Paradigm Life, Five Whys, Palette, IT STEP Academy, Turbine Workforce, Tailored AI and Seven Apps.

- **100,000 hours** checked and cut into training samples in under a week for one of the largest video data providers. Case study: https://videodb.io/blog/video-data-provider-100000-hours.md
- **3 weeks** from recordings to a first computer-use agent for a Big Four accounting firm's capability centers. Case study: https://videodb.io/blog/big-four-firm-gcc-agents.md
- **2×** the next pipeline's score on a 100-hour visual moment retrieval benchmark, when the clip must line up with the moment. Benchmark: https://videodb.io/blog/a-faster-path-to-the-next-policy.md

## Add VideoDB to the agent you already use

It takes about a minute. Then paste a link or upload a file and ask.

- **Claude** (web, desktop, mobile): add the plugin from the Claude plugin directory, https://videodb.io/claude, and authorize.
- **ChatGPT** (web, desktop, mobile, and Codex): add the plugin from ChatGPT, https://videodb.io/chatgpt, and log in to VideoDB once.
- **Claude Code**: `/plugin marketplace add video-db/skills`, `/plugin install videodb@videodb-skills`, then `/videodb setup`.
- **Cursor, Codex and any MCP client**: add `https://mcp.videodb.io/mcp` as a remote server (Streamable HTTP) and log in on first use. Steps per client: https://docs.videodb.io/pages/mcp/connect
- **Code**: `pip install videodb` or `npm install videodb`. Coding agents: `npx skills add video-db/skills`.

All install steps: https://videodb.io/agents.md

## What agents do with VideoDB

| Agent | Asked | What comes back |
| --- | --- | --- |
| Archive agents | "Cut three Shorts from this podcast." | `search_video` finds 3 moments; `reframe_video` cuts them as vertical clips, 9:16 |
| Computer-use agents | "What did the agent click before the error?" | `search_video` finds the click on Submit invoice at 14:02:11 |
| Physical agents | "Where did you last see the blue bottle?" | `search_rtstream` finds it on the kitchen counter at 09:41:07 |
| Research agents | "Two arms picking up objects in a kitchen." | `dataset_search` finds episodes from 9 datasets; `export_result_set` writes them as LeRobot files |
| Fleet agents | "A cyclist cutting across at dusk." | `search_collection` returns clips with the drive, time and camera |

Pages: archive agents https://videodb.io/archival-footage.md · computer use https://videodb.io/computer-use.md · physical agents https://videodb.io/blog/how-a-robot-remembers.md · research agents https://videodb.io/data-providers.md · fleet agents https://videodb.io/autonomous-vehicles.md

## Memory for agents and physical devices

Everything a camera, a screen or a file shows goes into an index as it arrives. A task gets back only the frames it needs.

- **Long-term**: the index of everything seen, written as it arrives, up to the live stream.
- **Short-term**: the frames a task needs, pulled from the index. Asked for the moment the robot comes back in color, the index points to 0:33.8 to 0:35.6, and those frames come back as a stream in 441 ms, the median for a five-second window.

Read how a robot remembers: https://videodb.io/blog/how-a-robot-remembers.md

## First on a 100-hour visual moment retrieval benchmark

VideoDB finds a matching clip for more queries than any other pipeline tested, on an open dataset with 11,369 queries. The dataset and the queries are open.

| Pipeline | Hit@10, any overlap | Hit@10, tIoU ≥ 0.5 | Hit@10, tIoU ≥ 0.75 |
| --- | --- | --- | --- |
| VideoDB | 40.39 | 20.49 | 12.83 |
| AWS Nova | 35.40 | 10.01 | 3.04 |
| Twelve Labs Marengo 3.5 | 32.87 | 9.61 | 3.37 |
| SigLIP2 | 24.96 | 6.22 | 1.90 |
| Qwen3-VL Embedding | 24.20 | 5.30 | 1.79 |
| Cosmos-Embed | 23.78 | 6.08 | 2.08 |
| InternVideo2 | 19.50 | 4.98 | 1.61 |
| PE-AV | 12.13 | 2.44 | 0.82 |

Hit@10 on 2,835 egocentric videos. tIoU is the time a returned clip shares with the moment, divided by the time either one covers. Under the any-overlap rule at top 50, AWS Nova (60.0%) and Twelve Labs (57.4%) find more than VideoDB (53.7%). Full results: https://videodb.io/blog/a-faster-path-to-the-next-policy.md

## From your files to your agent

Storage, search and streaming for video, in one system.

- **Files stay where they are**: buckets, cameras and screens, read in place.
- **Indexed by what happens**: any model describes every scene. Search runs on what it writes.
- **Frames on request**: about half a second from a question to the frames.

Runs on AWS, Google Cloud or Azure, or in your own cloud. SOC 2 Type II and GDPR: https://videodb.io/legal/security.md · https://videodb.io/legal/dpa.md

All products: https://videodb.io/products.md · The architecture: https://videodb.io/platform.md

## Open-source agents to start from

5,400+ stars on GitHub. Each repo runs on VideoDB and is free to fork: call.md (2,093 stars), Director (1,586), StreamRAG (505), Bloom (236). https://github.com/video-db

## What builders say

- "For a long time, I have been thinking what you are building needs to exist! Glad you are building it!" (Oliver Cameron, founder of Odyssey, ex-Voyage and Cruise)
- "VideoDB team, nice work! Would love to chat!" (Tom Mason, former CTO, Stability AI)
- "VideoDB is cool!" (Johnny Boufarhat, founder of Hopin)

## Recent research

- [How a robot remembers](https://videodb.io/blog/how-a-robot-remembers.md) (Oct 27, 2026): how a robot keeps an index of everything it has seen, and pulls back the few seconds a task needs.
- [VideoDB leads a 100-hour visual moment retrieval benchmark](https://videodb.io/blog/a-faster-path-to-the-next-policy.md) (Oct 6, 2026): VideoDB against AWS Nova, Twelve Labs and five open embedding models, on an open dataset with open queries.
- [Search over the Visual World](https://videodb.io/blog/search-over-the-visual-world.md) (Aug 8, 2026): the technical report. The right moment first 73.1% of the time, against 65.8% for a commercial video-native engine, over 9,834 queries on four public datasets.

All research: https://videodb.io/blog.md

## Try it on your own video

$20 of free credits and no credit card. Add it to Claude or ChatGPT, then paste a link and ask. [Add to Claude](https://videodb.io/claude) · [Add to ChatGPT](https://videodb.io/chatgpt)

About us, the team and the people backing VideoDB: https://videodb.io/team.md

To see is to know.

## Use VideoDB in your agent

VideoDB lets an AI agent watch, search and clip video.

- **Claude**: [Add VideoDB from the plugin directory](https://videodb.io/claude)
- **ChatGPT**: [Add VideoDB from the plugin directory](https://chatgpt.com/plugins/videodb)
- **Claude Code**: `claude mcp add --transport http videodb https://mcp.videodb.io/mcp`
- **Cursor or any MCP client**: Add the remote MCP server and log in once: `https://mcp.videodb.io/mcp`

Free to start: $20 of credits, no credit card. [Steps for every client](https://docs.videodb.io/pages/mcp/connect).
