# Computer use · VideoDB

> Build computer-use agents from how your team works.

The steps of most screen work are written down nowhere. VideoDB records the work as people do it and describes every step with the frames behind it, so you can build agents from it and check every run they make.

Canonical: https://videodb.io/computer-use · Talk to us: https://videodb.io/contact.md · Code: https://github.com/video-db/open-record-replay

The page opens with a short video of an open office floor, people working at rows of desks, and three handwritten notes: nobody wrote the steps down; "How does the team process an invoice?"; VideoDB records a few people doing it and writes down each step, then an agent runs them, and every run is on record.

## Record the work, step by step

The VideoDB Capture Agent records Windows and Mac desktops, or an agent's own virtual desktop. Each step becomes a description, a timestamp and the frames behind it.

- **Start**: by the person at the desk, or on a schedule.
- **Any model**: your VLM or ours names the app, the field, the click and the error, as text or JSON.
- **Local first**: an encrypted buffer on the machine, then an upload over HTTPS to your account or your cloud.
- **Afterwards**: once processed, a recording can move to cold storage or be deleted.

The page shows this on one recording: a person entering an invoice in a demo accounts app (invented app and data), recorded on Oct 31, 2026. VideoDB watched the 33-second recording with a vision model and returned one step for every 3 seconds, each as JSON with the step, the action, the screen element and the value: click the Harbor Supply Co. message in the Inbox list; click the Create bill button on the INV-20418.pdf attachment; type Harb into the Vendor field; select Harbor Supply Co. (V-1042) from the Vendor field; type INV-20418 into the Invoice number field; type PO-7731 into the PO number field; type 10/28/2026 into the Invoice date field; type 11/27/2026 into the Due date field; type 4,280.00 into the Amount (USD) field; click the Submit for approval button. The step that is playing is marked, and each step jumps to its moment.

How agents watch a screen: https://videodb.io/blog/openclaw-agent-camera.md

## From recordings to a first agent

The recordings give the agent its demonstrations, and give you the evidence to check it.

1. **Week 1, scope and record**: pick the task, agree what success looks like, and record a few people doing it.
2. **Weeks 2 and 3, build and check**: we write the steps from the recordings, then build the agent and check its runs against them.
3. **After launch, every run on record**: the agent records itself as it works, so any output traces back to the moment on screen behind it.

## A Big Four accounting firm sets up its capability centers with agents

Recordings of how experts do the work become SOPs, then computer-use agents. Every run the agents make is on record.

- **Customer**: one of the Big Four accounting firms.
- **Where**: its global capability centers (GCCs).
- **First workflows**: invoice processing, basic financial research and company valuation.
- **Data**: the region the firm picks; recordings return to its cold storage or are deleted.

The case study: https://videodb.io/blog/big-four-firm-gcc-agents.md

## Start from open code

Each repo runs on VideoDB and is free to fork.

| Repository | What it does |
| --- | --- |
| [open-record-replay](https://github.com/video-db/open-record-replay) | Compiles a recorded workflow into a reusable agent skill. |
| [openclaw-monitoring](https://github.com/video-db/openclaw-monitoring) | Records every agent run, so any output traces back to the moment on screen that produced it. |
| [videodb-capture-quickstart](https://github.com/video-db/videodb-capture-quickstart) | Captures screen, microphone and system audio into VideoDB. |
| [pair-programmer](https://github.com/video-db/pair-programmer) | Live screen context for Claude Code, Codex and Cursor. |
| [Bloom](https://github.com/video-db/bloom) | An agentic screen recorder. Recordings become inputs for AI. |

## About VideoDB

VideoDB is video infrastructure for agents and devices that see. 1,000+ developers and agents use it every week, and it has processed 300+ TB of video. It runs on AWS, Google Cloud or Azure, or in your own cloud, with a SOC 2 Type II report and a DPA for GDPR. https://videodb.io/legal/security.md

## Build your first computer-use agent with us

About three weeks from recordings to a first agent: one to scope and record, two to build and check it. [Talk to us](https://videodb.io/contact.md) · [Add to Claude](https://videodb.io/claude)