Real browsers your agent can open, read and photograph.
MCP Playwright lets your agent start Chromium or Firefox, open a page, read what is on it, click and type, and bring back screenshots, PDFs and video as evidence.
You can run several browsers side by side and ask all of them the same question at once, for example to see whether a page looks the same in both. Each browser is kept separate, so one browser's logins never leak into another.
What you gain over a generic browser tool.
A generic browser tool starts one disposable browser per run and pastes every image into the reply. The evidence is heavy, each browser engine is a separate setup, and nothing tells you which renderer drew the page.
Evidence by reference
You fetch only the evidence you need, and the conversation stays short. A pointer still uses some space, and the files last only for the session
- Typical alternative
- Image data pasted into the reply
- Here
- Screenshots, PDFs and video are saved and the reply only points to them
Several browsers, one setup
Each browser keeps its own website state and is easy to tell apart
- Typical alternative
- One tool setup per engine
- Here
- Chromium and Firefox work together, and you can see what is running at any time
One question, every browser
You compare engines without repeating work and see which one failed
- Typical alternative
- Repeating every step per engine
- Here
- A single request goes to all browsers in a group and comes back with one result for each
A chosen way of drawing pages
Pages are drawn the way you chose. A started browser does not prove every effect displays correctly
- Typical alternative
- Browser defaults, never reported
- Here
- Each browser starts with the drawing method you pick, and a start fails if a slow fallback shows up where fast hardware was expected
Sessions kept apart
The boundary is the conversation, not a separate machine. A shared mode deliberately shares ownership
- Typical alternative
- State shared between callers
- Here
- Browsers and their files belong to one conversation, and another conversation cannot read them
Browsers that clean up
You stop what you started instead of hunting for leftovers
- Typical alternative
- Orphaned browser processes left behind
- Here
- You start and stop each browser on purpose, and leftovers are stopped and their data removed when the conversation ends
Ready in the image, wired into your tools.
The service comes with the MAIT image and updates with it (mait-image). The browsers themselves arrive through a one-time install with AIDC, which keeps them in their own private runtime so they never touch the rest of your system. After that, nothing needs configuring.
You can also place it behind MCP Registry, so one entry in your agent's app reaches it.
With AIDE you get a local cluster with its own secure web address. It can point that address at the site you are building and keep live updates working, so your agent can drive a browser on the real address while your edits appear.
Where it sits in the MAIT toolkit.
It is the browser for your agent, and it pairs with the registry and with AIDE for live site work.
Where to read on.
- Get started, Use cases and Tools
- Playwright, the browser automation project the integration builds on
- playwright-mcp, the upstream MCP server whose page tools MCP Playwright exposes