Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

CUA Agent Kit

CVS Health's kit: an AI agent learns a browser or Windows workflow once, then Codex turns it into plain Playwright or pywinauto code.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.5/5

CUA Agent Kit gets the economics of screen automation right by paying for AI once and running code after, but it is an early starter kit with sharp edges around credentials and an unauthenticated Windows server.

What is CUA Agent Kit?

CUA Agent Kit comes from CVS Health and splits screen automation into two jobs. In discovery, an OpenAI computer-use agent (the default model is gpt-6-sol, with 40 turns) follows a workflow written in plain language in a browser or on a remote Windows desktop, recording screenshots, actions, targets and, in the browser, a Playwright trace. In codification, a Codex skill reads the successful recordings and writes ordinary code: Playwright with semantic locators for websites, pywinauto with UI Automation and Win32 controls for Windows apps. Repeat runs use that code with no model calls and no API key.

It is a Python 3.12 project managed with uv, with a local sample intake app, optional PostgreSQL audit records and Azure Blob Storage for artifacts, and Docker Compose files. The repository was created on 30 October 2026 and has 19 stars, an Apache-2.0 licence, 11 commits and no releases; the latest commit is from 12 November 2026.

Is CUA Agent Kit worth it?

Computer-use agents can work out a clicky back-office task from a paragraph of instructions, yet asking a model to reason through the same clicks every morning adds cost and variance to work a script can do. This kit treats the agent as a discovery tool. You rerun discovery until the workflow succeeds consistently, ask Codex to codify those runs, and from then on `cua-agent-run` replays deterministic Playwright or pywinauto code. The project credits Kelsey Hightower's Zero Token Architecture talk at PlatformCon 2026 for the idea.

The documentation is unusually honest about where the method stops. Its approach guide says Citrix-delivered apps that expose only pixels have no controls to codify, so they stay AI-driven on every run, and that a hybrid runner mixing code with model steps is not provided. It also warns that an agent finishing a run does not prove the task succeeded. When a site changes, the codified script breaks like any RPA script; the kit's answer is to have Codex read the new traces and update the code, with an engineer reviewing.

The catch: discovery sends your workflow prompt to the model endpoint, including the login credentials injected from `WORKFLOW_ACCOUNT_JSON`, along with screenshots of whatever is on screen. Recordings can hold cookies, session tokens, network responses and DOM snapshots, and redaction is described as best effort. Enabling the storage backend uploads the full run folder unredacted. On Windows, the host server can control the desktop and run commands, and the setup scripts add no authentication or firewall rules, so it belongs on a locked-down private network only.

Our take: this is a well-reasoned blueprint from a large employer rather than a finished product, with 19 stars and 11 commits; a team with Python engineers can lift the pattern, while anyone wanting a supported RPA tool should look elsewhere.

Where it earns its place

  • An operations engineer at a healthcare provider

    Automate a daily referral-entry task in a web portal with no API

    Describe the steps in `prompt.py`, run `cua-agent-discover` against a test account until it succeeds reliably, then have Codex codify it; the daily run is a Playwright script with no API spend. Use a test login, since credentials reach the model during discovery.

  • An IT automation team supporting a legacy Windows desktop app

    Replace a fragile coordinate-based macro with control-aware code

    Install the host server on a Windows VM with `install.ps1`, discover the workflow, and codify it into pywinauto steps that use UI Automation controls. Check representative screens first: an accessible main window does not mean every dialog is.

  • A platform team assessing computer-use agents

    Compare per-run AI automation with discover-once codified scripts

    Run the included Care Intake sample in both modes and record cost and failures. The kit's own guidance is that apps exposing only pixels, such as some Citrix apps, must stay AI-driven every run.

Before you commit

Setup
Clone the repository, `uv sync`, set `OPENAI_API_KEY`, and run `uv run playwright install chromium`; the sample app starts with `uv run python sample_app/app.py`. Windows needs a VM prepared with `windows-host/install.ps1` and `start.ps1`, kept unlocked during runs.
Integrations
OpenAI computer-use models (or a compatible endpoint via `OPENAI_BASE_URL`), the OpenAI Agents SDK, Codex skills, Playwright, pywinauto, PostgreSQL with Flyway migrations, Azure Blob Storage and Docker Compose.
Data and privacy
Discovery sends prompts, injected credentials, screenshots and action history to the model endpoint. Artifacts stay local by default under `artifacts/`; recordings can contain credentials, cookies, tokens and network data, and storage uploads are not redacted. The Windows server has no authentication.
Cost at scale
Model spend falls on discovery and on re-discovery when a workflow changes; codified runs make no model calls. No per-run cost figures are published.
Limits
Discovery is capped at 40 turns by default. No ready-made hybrid runner. Apps without accessible controls cannot be codified. Codified scripts break when the target UI changes and must be re-codified.

What’s good

  • Codified runs need no model calls or API key
  • Playwright locators and pywinauto controls instead of screen coordinates
  • Clear documentation of security risks and method limits
  • Optional PostgreSQL audit trail and Azure Blob artifact storage

What’s not

  • Discovery sends credentials and screenshots to the model
  • Recordings may contain cookies and session tokens
  • Windows host server has no authentication
  • Very young: 19 stars, 11 commits, no releases

Best for

Engineering teams in operations-heavy organisations who want to automate repetitive browser or Windows screen work and keep the repeat runs free of model costs.

CUA Agent Kit alternatives

CUA Agent Kit: common questions

Is CUA Agent Kit worth it?
CUA Agent Kit gets the economics of screen automation right by paying for AI once and running code after, but it is an early starter kit with sharp edges around credentials and an unauthenticated Windows server. We rate it 3.5 out of 5.
How much does CUA Agent Kit cost?
Free and open source (Apache-2.0). Discovery runs bill to your OpenAI API key; codified runs make no model calls.
What are the downsides of CUA Agent Kit?
Discovery sends credentials and screenshots to the model; recordings may contain cookies and session tokens; windows host server has no authentication; very young: 19 stars, 11 commits, no releases.
Who is CUA Agent Kit best for?
Engineering teams in operations-heavy organisations who want to automate repetitive browser or Windows screen work and keep the repeat runs free of model costs.
How much does CUA Agent Kit cost for a team or in production?
Model spend falls on discovery and on re-discovery when a workflow changes; codified runs make no model calls. No per-run cost figures are published.
Is CUA Agent Kit safe for company data?
Discovery sends prompts, injected credentials, screenshots and action history to the model endpoint. Artifacts stay local by default under `artifacts/`; recordings can contain credentials, cookies, tokens and network data, and storage uploads are not redacted. The Windows server has no authentication.
What are the best CUA Agent Kit alternatives?
The closest alternatives we have reviewed are Skyvern, Browser Use, Stagehand and UFO.

Answers checked Nov 20, 2026.