[go: up one dir, main page]

Products

Agentic video primitives

Four layers behind one API, from the first frame to your agent. Use the ones you need, in our cloud or yours.

  • About 0.5 sfrom a question to the frames that answer it
  • 100,000 hourschecked and cut into samples in under a week
  • 300+ TBof video processed
  • 1,000+developers and agents use VideoDB every week

Primitives

Files, buckets, cameras and screens go through one pipeline. Ask for any frame or window and get exactly that.

See the architecture
Ingest
RTSP cameras, screens, microphones, files, links, S3, GCS and Azure buckets
Transcoding
In-house and distributed. Seeking costs the same anywhere in the file.
Frames
Any frame or window on request. Tiling packs a window into one image for a model.

Understanding

Any model, open-weight, proprietary or your own, writes what it sees into your own JSON fields.

understand_video

Homepage video · 37.8 s
{
  "segmentation": { "type": "time", "seconds": 2 },
  "analyzers": [{
    "type": "vlm",
    "name": "scene",
    "sampling": {
      "strategy": "uniform",
      "frame_count": 3
    },
    "config": {
      "model": "pro",
      "prompt": "Describe what is visible…"
    }
  }]
}
  1. Split into 19 segments of 2 s
  2. Sampled 3 frames from each segment
  3. Wrote one scene record per segment

Done: 19 records, one for every 2 seconds of the homepage video

Scene records

Complete

  1. 0:00–0:02

    A small round robot in the foreground looks out over a misty mountain valley with a winding river and a cluster of illuminated village houses below…

    scene
  2. 0:06–0:08

    A color landscape of snow-capped mountains and a misty valley, partially obscured by a semi-transparent, regular grid of colored dots…

    scene
  3. 0:12–0:14

    Abstract, densely layered contour-like lines and swirls that look like a topographic map or fingerprint patterns…

    scene
  4. 0:22–0:24

    Closely spaced curved contour lines that form a jagged, fractal edge on the left and broad sweeping arcs to the right…

    scene

On WGO-Bench, a VLM pipeline reaches 28.9% semantic F1 at IoU 0.5, against a 19.9% public baseline.

Read the annotation benchmark

Indexes

Ask in plain language across a whole collection, then filter and count. Every result is a clip sized to the question.

Found 3 scenes with a robot in them.

Scene index

3 found

Best match

market-robot.mp4 · 0:00–0:27.3

A robot at a workbench in a neon market

  1. white-robot.mp4 · 0:00–0:07.6

    A white robot dancing

  2. stage-robot.mp4 · 0:05–0:24

    A robot dancing on a lit stage

VideoDB puts the right moment first 73.1% of the time, against 65.8% for a commercial video-native engine.

Read the technical report

Delivery

Every answer can be a playable stream, an event or a file. Nothing is rendered ahead.

Read the docs (opens in a new tab)

python

q = "village houses with their lights on"
res = video.semantic_search(
    query=q, top_k=4,
    index_ids=[idx.index_id])

# every result is a playable stream
stream = res.compile()
  1. MCPTools for any MCP client
  2. SDKPython and Node
  3. WebSocketEvents with the clip
  4. WebhookWhen a job is done
  5. ExportLeRobot, MCAP, Parquet

Run in our cloud or yours

The files can stay in your bucket, in the region you choose. Same SDK and MCP server either way.

Security

Runs on any cloud

AWS, Google Cloud and Azure, or your own cloud.

Security and compliance

SOC 2 Type II, GDPR and ISO 27001 compliant.

  • SOC 2 Type II
  • GDPR
  • ISO 27001

Quickstart

Get your API key at console.videodb.io. Then upload a file, understand it, index it, search it and play the result.

pip install videodb
quickstart.py
from videodb import connect

# uses your API key from console.videodb.io
conn = connect()
coll = conn.get_collection()
video = coll.upload(url="s3://lab/run-0412.mp4")

u = video.understand(analyzers=[{"type": "vlm", "name": "scene",
    "config": {"prompt": "What happens, and with what result?"}}])
u.wait_until_complete()
idx = video.index(name="scene", source=u.get_analyzer("scene"))

res = video.semantic_search(query="the gripper drops the bottle",
    index_ids=[idx.index_id], top_k=5)
for shot in res.get_shots():
    print(shot.start, shot.end, shot.metadata)

# the result, as a playable stream
stream = res.compile()

https://videodb.io/products.mdOpen the file

## 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).