Playwright: how AI agents work with software that has no API
TL;DR
Playwright is a free tool that lets an AI agent log in to software and use it like you do. It is the answer when a product has no API, or has one that is locked down. The catch is the terms of service. Some vendors ban agents and will cut your access if they spot one. Check the terms first, then work at a human pace.
What if the software has no API?
Most business software was built for people. Screens, menus, buttons, a login.
Agents do their best work through an API or an MCP connection, which is a clean channel straight to the data. But plenty of the software you rely on has neither. Or it has an API that exposes a fraction of what you can do on screen.
So you hit a wall. The task is simple, the agent is capable, and there is no door.
Playwright is the door.
What is Playwright?
Playwright is a free, downloadable tool that drives a real web browser from code. Point an agent at it and the agent can open the site, log in with your credentials, click the buttons, fill in the forms and read what comes back.
It does what you do. It just does it without you sitting there.
That means you are no longer restricted by:
- whether the software has an API
- whether the software has an MCP server
- whether the vendor ever planned for an agent to use it
If a human can do it in a browser, an agent can be taught to do it through Playwright.
How does this differ from computer use?
Most of the big AI models now ship with something called computer use. Claude has it. Grok has it. The model looks at a screen, works out what is on it and moves a mouse and keyboard.
It is a useful capability, and it is the same idea. The model acts like a person at a computer.
Playwright is how I run it for routine work. It is free, it lives inside the browser, and it repeats the same steps the same way every time. The model supplies the judgement. Playwright supplies the hands. Together they log in, click through and pull the data out.
Why is this ideal for restricted software?
Three situations come up again and again:
- No API exists. The only way in is the front door.
- The API is restricted. It covers the basics but not the report, export or setting you actually need.
- The data sits behind a login. No connection to build, just a screen to read.
In all three, you stop asking "does this software support agents?" and start asking "can a person do this in a browser?" The second question has a much better hit rate.
See where AI fits in your business. Free.
A 45-minute audit. We map the highest-value automations and what they're worth in time and money. No pitch, no pressure.
What is the danger?
Here is the part people skip.
Some software has a basic premise in its terms of service: you are not allowed to use AI on it. When your agent logs in, the system recognises it as automated and blocks it. The terms can say, in plain words, that they will remove your access if they catch it. You lose the software. That is the risk if you are not careful.
So the order matters.
- Check whether the vendor allows agents. Read the terms. Search for "automated", "bot", "scraping" and "AI". If it is unclear, email them and ask.
- If they allow it, go ahead. Use your own account and your own data.
- If they ban it, stop. Ask for an API, ask for an exemption, or pick different software.
Do not treat the third case as a puzzle to solve. A banned use stays banned however well the agent behaves.
Why should the agent work at a human pace?
Even where agents are allowed, speed is a problem.
When you first watch an agent work, it is startling. It knows exactly where to click. It never searches for a menu, never rereads a label, never pauses. It does in seconds what takes you twenty minutes.
You are going in at a computer's pace. And that pace looks nothing like a customer. Systems built for people react badly to it. You trip rate limits, you fill their logs with noise, and you risk a block even on software that permits you to be there.
So tell the agent to take its time:
- wait for each page to finish loading before it acts
- add short, uneven pauses between steps
- run big jobs in small batches, spread across the day or the week
- stop and report if anything looks different from what it expected
This is courtesy and accuracy, not a loophole. A slower agent also reads the page properly instead of clicking on something that has not finished loading.
What does headless mean?
You are going to hear the word "headless" a lot more over the next few years.
All it means is software that an agent can use directly. A human needs an interface, so software has one. An agent does not need the screens and buttons. It needs a route to the data and a way to take actions.
Playwright sits in the gap. It is the bridge between a human interface and an agent when there is no backend open to it. It is a workaround, and a good one, for software that has not caught up.
Where is this heading?
Software is splitting in two.
A lot of legacy products have not made the jump to agentic. They have an interface for humans and nothing for agents. As agents get smarter and more businesses run on them, the products with no agent side will lose ground. If an agent cannot get at the data, it will use the product that lets it.
Within two to three years, maybe five for the ones with long contracts to hide behind, most new software will ship with two front doors. One for people. One for agents.
If you are cursing a tool today because it will not let an agent in, that is your early warning. Ask the vendor what their agent plan is. Their answer tells you how long to stay.
What should you do this week?
- Make a list of the software you use that has no API or a weak one.
- Check each one's terms for automated access.
- Pick one task in a permitted tool that eats an hour a week.
- Build it with Playwright and run it at a human pace.
Start with your worst task. If the software has no door for your agent, you do not have to wait for the vendor to build one.
Live with passion & AI,
Brett
Need an AI operator inside your team?
Place a Chief AI Officer, an AI Officer, or embed an Anaboo Forward Deployed Engineer for 3–6 months.
Frequently asked questions
What is Playwright?
+
Playwright is a free, open source tool that controls a real web browser from code. An AI agent can use it to log in to a website, click buttons, fill in forms and read the screen, exactly as a person would.
When should I use Playwright with an AI agent?
+
Use it when the software you need has no API, when the API is restricted or missing the feature you need, or when the data only exists behind a login screen. If a proper API or an MCP connection exists and does the job, use that first.
Is it allowed to use an AI agent to log in to software?
+
That depends on the software's terms of service. Some vendors ban automated access outright and can close your account if they detect it. Read the terms before you start, and ask the vendor if the terms are unclear. Slowing the agent down does not make a banned use allowed.
What does headless mean in AI?
+
Headless means software that an agent can use directly, without a human interface in the way. A human needs buttons and screens. An agent only needs a route to the data and the actions.
What is computer use?
+
Computer use is a built-in ability in several AI models, including Claude and Grok, to look at a screen and operate it with a mouse and keyboard. Playwright does the same job inside a browser, and it is faster and more repeatable for routine tasks.
Why does the pace of the agent matter?
+
An agent moves at machine speed. It knows where every button is and never hesitates. Many systems read that speed as abuse and respond with rate limits or blocks. Pacing the work like a person protects your account and gives the page time to load, so the agent reads what is really there.
Brett is a four-time founder (Darra Tyres, Gladfish, EzyTrac, Anaboo) and the operator behind AIOS, Anaboo's AI Operating System. He writes from inside the build, installing AI in his own businesses first and reporting back what actually moves the numbers. Based between Singapore, the UK and Australia.
