weblab
Attaches to an already running Electron app as a browser target, enabling agents to interact with and inspect Electron-based apps while leaving the app as it was found.
Runs browser sessions in Firefox alongside other engines, allowing cross-browser testing and exploration of web apps with the same step-based actions, assertions, screenshots, and logs.
Drives Google Chrome as the default real browser target, enabling AI agents to open sessions, run browser steps, read page state, take screenshots, log console/network activity, execute JavaScript or CDP commands, and test or explore web apps.
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
weblab
weblab is an MCP server that drives real browsers, so AI agents can test and explore web apps. An agent opens named sessions (each a browser of its own pointing at an address), runs steps on them, and reads back what happened as text and screenshots.
Install
curl -fsSL https://raw.githubusercontent.com/dittofleet/weblab/main/install.sh | shOn Windows, in PowerShell:
irm https://raw.githubusercontent.com/dittofleet/weblab/main/install.ps1 | iexThis puts one binary at ~/.local/bin/weblab, or %LOCALAPPDATA%\Programs\weblab\weblab.exe on Windows (set WEBLAB_INSTALL_DIR to change that). weblab runs on macOS, on Apple silicon and Intel, on Linux on x86_64, and on Windows on x64. It drives a browser already on the machine: Google Chrome by default, or another Chromium browser. The install also fetches Playwright's small build of ffmpeg, which the video step records through.
On Linux
Without a display, as on a server, headed is refused, and an Electron app needs a virtual one, which Sessions shows how to give it. screen shots work on macOS only.
On Windows
Sessions, servers, recording and attaching work as elsewhere. screen shots work on macOS only. Windows can't ask a console program to close, so a server weblab started is stopped outright rather than given time to shut down.
To update, run weblab update. A running weblab looks for a newer release once a day, and says so in the reply to new when there is one.
Related MCP server: Playwright MCP Server
Register it with an MCP client
weblab speaks MCP over stdio, so the client starts it:
claude mcp add weblab -- weblab
codex mcp add weblab -- weblabIn a client's JSON config, the server entry is { "command": "weblab" }. Sessions last as long as that weblab process does: when the client goes, every session ends, and a server weblab started for them stops unless another weblab is still using it.
The tools
Tool | What it does |
| Opens a session: a browser of its own, with a name, pointing at an address. Runs the |
| Runs steps on a session, in order, stopping at the first that fails. Replies with each step's outcome, what it handed back, and its screenshots. |
| Ends one session, or all of them. Writes a video still recording and traces, and stops a server nobody else is using. |
| The sessions that are open, where each one is, and whether steps are running on it. |
| The reference pages below, readable from inside the client, whole or a section at a time: |
An example
new
{ "address": 5173, "start": "pnpm vite --port 5173", "path": "/settings" }session main at http://localhost:5173 (server started by weblab: pnpm vite --port 5173)
files $TMPDIR/weblab/shop-20261003-101600
started the server at http://localhost:5173 (pnpm vite --port 5173 in ~/code/shop); it stops when the last session on it ends
ok 1 goto (1840ms)
url http://localhost:5173/settings
title Settings
console since the last reply:
[console.log] app mounted (http://localhost:5173/src/main.tsx)run
{
"steps": [
{ "click": { "role": "button", "name": "Appearance" } },
{ "click": { "label": "Dark" } },
{ "expect": { "js": "document.documentElement.classList.contains('dark')" } },
{ "js": "localStorage.getItem('theme')" },
{ "shot": "appearance" }
]
}ok 1 click (41ms)
ok 2 click (37ms)
ok 3 expect (12ms)
ok 4 js (3ms)
"dark"
ok 5 shot (95ms)
shot $TMPDIR/weblab/shop-20261003-101600/shots/main-appearance.pngThe screenshot also comes back as an image in the same reply. A step that fails is shown as FAIL 3 expect: ... with a screenshot of the page at that moment, and the session stays open for the next call.
What it can do
Start the app with the command it is given when nothing answers at its address, share it among every session pointing there, and stop it when the last one ends. A server it didn't start is used as it is and left running. It never guesses how a project is run.
Run any number of sessions at once: two users, copies of an app on several ports, Chrome, Edge, WebKit and Firefox side by side.
Join a browser or Electron app that is already running, and leave it as it was found. Run code in an Electron app's main process too.
Read the page as an accessibility tree with refs to act on, as text, or as HTML.
Act and check: click, type, fill, drag, upload, and
expectwaits for text, elements, URLs, requests, console messages or any JavaScript.Shape what the page sees: fake API responses, colour scheme, viewport, locale, a saved sign-in, a slow network.
See a React app the way React DevTools does: its components, where each one's code is, and what rendered and why. Change props, state and boundaries on the spot.
Run JavaScript in the page, Playwright code against it, or raw Chrome DevTools Protocol commands, for anything the built-in steps don't cover.
Keep screenshots, compare one with an earlier one, capture native menus from the real screen, record video and Playwright traces, and log the console and network.
Documentation
Page | What's in it |
Every argument of | |
Every step and its options, targeting elements, refs, | |
More than one session: two users, copies of an app, running browsers and Electron apps, other engines, saved sign-ins. | |
| |
Short answers to common tasks, as the steps to run. |
Development
bun install
bun test # end-to-end tests, in the installed Chrome
bun run build # dist/weblab, a single binarysrc/main.ts is the entry point and src/mcp.ts defines the tools. Each step lives in src/steps/, grouped as Steps lists them. A release is a vX.Y.Z tag pushed on a commit on main, and nothing else: the release workflow builds the binaries, stamps the tag in as their version, and publishes them. No file in the repository holds the version, so there is nothing to bump.
The API is designed for how it reads now, not for what an earlier version took. Agents learn it from the tools' own descriptions and the docs each time they connect, so nothing depends on an old name the way a script would. When a clearer name or shape turns up, change it.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Control a real browser from AI agents: navigate, click, type, screenshot, read the page.
Give AI agents a real cloud Chrome: open sessions, navigate, read pages, click, type and screenshot.
Let your AI agent operate any website through a real, logged-in browser session.
Write, run and check natural-language browser tests of your web app from your AI assistant.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables AI coding agents to autonomously interact with and test web applications in a real browser, providing DOM/Accessibility tree extraction, runtime telemetry, screenshot capture, and Markdown test reports.9 npm1MIT
- AlicenseNot gradedqualityDmaintenanceEnables LLMs to drive a real browser for web scraping, page interaction, screenshots, JavaScript execution, automated test generation, and responsive device emulation through natural language tool calls.MIT
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to perform end-to-end web testing by controlling a real browser through tools for navigation, interaction, state inspection, screenshots, and script evaluation.MIT
- AlicenseNot gradedqualityBmaintenanceEnables LLM agents to drive a selector-free browser through natural language, so they can debug styling and console errors, map and document web apps, and generate or run end-to-end and exploratory tests without screenshots.908 npm88MIT