Infrastructure for AI agents

Where AI agents see video.

Giizu is the video layer for the agent economy: footage that’s already understood, compute right beside it, and a preview that streams live.

Request early access How it works

Or write to [email protected]

Move your cursor: it paints.

An interactive example. Type a request such as “find the drone shot” or “make it moody” and four sample shots respond.

ready · 42 clips indexed
try:

Type your own request or tap one. An illustrative demo of what agents do on Giizu.

The journey of a frame

01 / 05

散

Hours of footage, scattered.

More than anyone has time to watch, let alone an agent.

解

Giizu understands every frame.

Transcribed, described and logged the moment it lands.

探

Agents find what matters.

They read the layers, not the raw video.

編

Then cut it, live.

Their own code runs beside the footage, and the preview streams back.

完

Rendered on the originals.

The same code, re-run on the camera files.

SCROLL

The journey of a frame

  1. 散
    Hours of footage, scattered.More than anyone has time to watch, let alone an agent.
  2. 解
    Giizu understands every frame.Transcribed, described and logged the moment it lands.
  3. 探
    Agents find what matters.They read the layers, not the raw video.
  4. 編
    Then cut it, live.Their own code runs beside the footage, and the preview streams back.
  5. 完
    Rendered on the originals.The same code, re-run on the camera files.
一The gap

Video is the biggest kind of data agents still can’t touch.

They can’t see what’s inside footage, and they can’t work on it where it’s stored. So today they download first, and guess.

75%

of mobile data traffic is video, the largest category online.

4.4M

tokens for an agent to look at one hour of 720p, more than four context windows.

22 h

and about $90 in egress just to move 1 TB to where the agent’s code runs.

Getting an agent to work on 1 TB of footage

Download, then edit
~22 h · ~$90 egress
Giizu
0 bytes moved

Ericsson Mobility Report (2025); Anthropic image-token formula at 1 fps; AWS S3 egress at $0.09/GB over a 100 Mbps line.

二The flow

Upload once. Let agents do the rest.

Footage is understood the moment it lands. Agents read it, run their own code on it, and you watch the result stream back.

Understood on arrival

Transcripts, descriptions, faces, shots, contact sheets and a proxy, organised into layers.

Agents work in place

Claude, Codex or your own agent reads the layers and runs its code right next to the footage.

Preview streams live

Every change plays back in seconds. When it’s right, the same code renders on the originals.

三The context

Agents read a brief, then zoom in to the frame.

Layer 0

Project brief

A summary of everything you shot: people, places and story beats.

~10ktokens to understand the whole project

  • Layer 0 · Project briefAbout 10k tokens covering everything you shot.
  • Layer 1 · ScenesEach scene summarised: who, what, where it sits in the story.
  • Layer 2 · ShotsEvery shot logged with transcript, faces and a description.
  • Layer 3 · Contact sheetsThumbnails to check a look before touching pixels.
  • Layer 4 · Frames and proxyExact pixels and the 720p proxy, only when needed.
四The primitives

Four calls any agent can compose.

Connect over MCP or our API. Hover a primitive to see it in the script.

read

Zoom from the project brief down to a single frame.

run

Execute the agent’s own code beside the footage.

preview

Stream every change back in seconds.

render

Re-run the same code on the camera originals.

# an agent cutting a launch film (illustrative)gz.read("launch-film", depth="brief")gz.read(query="founder laughing, golden hour")gz.run("python grade.py --lut film.cube")gz.run("python teaser.py --length 15")gz.preview(timeline)gz.render(timeline, source="originals")→ master ready
五The bet

We sell the kitchen, not the menu.

Agents are strong programmers now. They want raw tools and a place to run them, not a fixed list of operations.

Yesterday · the menu

trim()
add_text()
apply_filter()

Fixed operations built for agents that needed guardrails. New features wait on someone else’s roadmap.

Giizu · the kitchen

FFmpeg · OpenCV
segmentation models
OpenColorIO · your code

Give an agent the tools and the footage in one place, and it builds the feature itself.

六Questions

Frequently asked questions

What is Giizu?

Giizu is the infrastructure layer where AI agents see, edit and ship video. Footage is understood on upload, agents run their own code right beside it, and every change streams back as a live preview.

How do AI agents connect to Giizu?

Agents connect over MCP or the Giizu API. It works with Claude, Codex or your own agent, which reads layered context about your footage and calls four primitives: read, run, preview and render.

Do I have to download footage to edit it?

No. You upload footage once. Agents work on a 720p proxy next to where it is stored, preview changes as a stream, and the same code renders the final cut on the camera originals.

Which tools can an agent use on Giizu?

Agents can run their own code with FFmpeg, OpenCV, segmentation models and OpenColorIO preinstalled, instead of being limited to a fixed menu of video operations.

Why not send video straight to a language model?

Looking at one hour of 720p video at one frame per second costs an agent about 4.4 million tokens, more than four context windows. Giizu gives agents a project brief of roughly 10,000 tokens and lets them zoom in to exact frames only when needed.

Is Giizu available today?

Giizu is onboarding a small group of early teams building with video and agents. Write to [email protected] to request early access.

Early access

When agents make video, they’ll make it on Giizu.

We’re onboarding a small group of teams building with video and agents.

Tell us three things

  1. 壱What you’re building
  2. 弐Roughly how much footage you have
  3. 参Which agent you use: Claude, Codex or your own
Write to us