// levr browser · a browser mcp for your agents
Your agent says it tested. Now you can see it, and prove it.
Levr Browser turns your coding agent into a testing agent, in your own browser. It clicks through your app the way a person would, and records everything it does on the issue.
Chrome, Edge and other Chromium browsers · Works with Claude Code, Cursor, Codex and any MCP client
⏺ Testing checkout in the Levr tab group before I mark ENG-512 done.
⏺ browser_navigate(localhost:3000/checkout)
⎿ Levr tab group · tab 1
⏺ browser_fill_form(card, expiry, zip)
⎿ 3 fields filled
⏺ browser_click(“Pay $42.00”)
⎿ clicked
⏺ browser_get_console_logs(level: error)
⎿ TypeError: total is undefined
⏺ browser_screenshot(attach_to: ENG-512)
⎿ saved · checkout-error.png
⏺ browser_save_har(attach_to: ENG-512)
⎿ saved · 38 requests, 1 failed
⏺ Checkout is broken: the total is never set when a coupon is removed. Screenshot and network log are on ENG-512. Fixing it now.
It’s not “done” until it’s tested. In a real browser.
Unit tests pass and the page still won’t load. An agent that can open the app, click through it and read the console finds that out before you do. Levr Browser is the agentic browser for that job: your browser, your logins, and a record of exactly what the agent did.
// the testing automation never did
Hand your agent the manual testing. Every step lands on the issue.
Most testing was never automated, because it was never worth scripting: exploratory sessions, ad hoc checks, reproducing a bug, eyeballing a fix. So it stayed a person clicking around, and what they saw lived in their head or a screenshot in chat. Levr Browser lets your agent do those chores, and records everything it does on the issue, automatically. That’s what no other test automation does.
Chores your agent can take now
- Exploratory testingPoke at a new feature the way a curious user would, and report what broke.
- Ad hoc checks“Does the coupon still apply after a refresh?” Answered in a minute, with proof.
- Reproducing a bug reportFollow the steps from the issue, confirm the bug, and capture the failing request.
- Verifying a fix by handWalk the flow again after the change, instead of trusting a green unit test.
- Phone and tablet spot checksEmulate the device and look at the page the way your users will.
- Logged-in and real-data flowsIt uses your own browser, so the accounts and sessions a fresh test browser lacks are already there.
Recorded on the issue, without asking
- Every action, step by step, in the action trace
- Screenshots of what it saw
- A video recording of the session
- Console errors and warnings
- Every network request, as a HAR file, credentials redacted
- Your own clicks too, if you turn on input recording
And when a check should run every release, the same session becomes a regression test: export the trace and replay it. See how.
// how it connects
Three steps, and nothing listens on your machine.
Your agent calls the levr-browser MCP server on the Levr API. Levr routes each request to your signed-in browser. No local server, no native host, no port to open.
- 1Install from the Chrome Web StoreWorks in Chrome, Edge and other Chromium browsers, and updates itself.
- 2Sign in to LevrFrom the side panel. Your browser shows up as connected in your workspace.
- 3Connect your agentAdd the levr-browser MCP server to Claude Code, Cursor, Codex or any MCP client. Settings › Connect an agent prints the exact setup.
# Claude Code
$ claude mcp add --transport http levr-browser https://ai.levr.one/api/v1/mcp/browser
# Cursor, Codex, any MCP client: add an HTTP server
levr-browser → https://ai.levr.one/api/v1/mcp/browser
✓ sign in once in the browser when asked · no token to paste
// what your agent can do
A browser agent that tests like you would.
Around fifty browser tools, all in the tabs you share. These are the ones you’ll notice.
Drive the page
Navigate, open tabs and windows, click, type, press keys, hover, scroll, drag, fill forms, handle dialogs and upload test files.
- browser_navigate
- browser_click
- browser_fill_form
Read what happened
An accessibility snapshot of the page, its content, and the console and network logs, so the agent debugs from facts.
- browser_snapshot
- browser_get_console_logs
- browser_get_network_logs
Keep the evidence
Screenshots, full-resolution tab recordings up to five minutes, HAR files, console captures and performance traces, saved to your workspace.
- browser_screenshot
- browser_start_recording
- browser_save_har
Test on any screen size
Emulate a device or a viewport and see the page the way a phone or a tablet does.
- browser_emulate
- browser_select_device
Turn a session into a test
Mark the steps that matter, export the action trace as a step-based Levr test, then replay it and record each step’s result on a run.
- browser_mark
- browser_export_trace_as_test
- browser_replay
Show it, then hand it over
Record your own clicks and typing as a trace (off by default), so the agent can pick up exactly where you left off.
- capture_user_input
// evidence on the record
What the agent saw, not what it says it saw.
Screenshots, recordings, HAR files, console logs and performance traces save straight to your workspace and attach to the issue, test or run result they belong to. Open any of them in Levr: the recording plays, the HAR has a request list, the trace opens in a viewer.
When a gate needs a hands-on check, this is how an agent shows what it observed. Read about Levr Gates.
- recordingcheckout-flow.webm0:42 · 1920×1080
- screenshotcheckout-error.pngfull page
- harcheckout.har38 requests · 1 failed (500)
- consoleconsole.log1 error · 3 warnings
- tracecheckout-trace.jsonperformance · 4.1 s
// traces become tests
Explore once. Replay forever.
Every session is an action trace you can watch step by step. Export it as a step-based Levr test, replay it later, and each step’s result is written to the test run. Exploratory testing that leaves a regression test behind.
Want to show the agent the way? Turn on input recording, click through the flow yourself, and hand the trace over.
- 1passOpen /checkout
- 2passFill card, expiry and zip
- 3passRemove the SAVE10 coupon
- 4passTotal shows $42.00
- 5passPay succeeds
// your browser stays yours
Built so you can hand it your real browser.
- Agents work only in tabs in the “Levr” tab group: tabs they open, or tabs you choose to share. Your other tabs, history and bookmarks are never read.
- Nothing listens on your computer. Agents reach your browser only through your Levr workspace, and need your sign-in.
- Stop all, in the side panel, stops every agent at once. Agents can’t resume on their own.
- Password, card and one-time-code fields are recorded as their length only, never their value.
- Captures you don’t file are deleted after an hour by default. Filed ones stay with their issue, test or result.
- No analytics, no ads, no selling data, and no remote code.
The details are in the Levr Browser privacy policy.
// your browser vs a fresh one
Why not a headless browser MCP?
By default, Playwright MCP and Chrome DevTools MCP give an agent a browser of its own, and they’re great for that. Levr Browser is for when the test needs your browser and the result needs to be kept.
| Question | Levr Browser | A headless browser MCP |
|---|---|---|
| Which browser | Your own Chrome or Edge, with your logins and extensions | A fresh, separate browser the agent launches |
| What the agent can touch | Only tabs in the Levr tab group | Everything in its own browser |
| Where the evidence goes | Attached to the issue, test or run it proves | Files on the machine the agent ran on |
| Watching and stopping it | You see every action in the side panel. Stop all ends it. | Depends on the setup |
| Turning a session into a test | Export the trace as a Levr test and replay it | Write one yourself |
// faq
Levr Browser, answered.
What is Levr Browser?
A Chrome extension that lets your AI coding agents test your web app in real browser tabs: open pages, click, type, fill forms, take screenshots, and read console and network logs. Screenshots, recordings and results are recorded in your Levr issues, tests and runs, so you see what the agent actually tested.
What is a browser MCP?
An MCP server that gives an AI agent browser tools: navigate, click, type, read the page, take a screenshot. Levr Browser is a browser MCP whose browser is your own: the agent calls the levr-browser MCP server on the Levr API, and Levr routes each request to your signed-in extension.
Is it for manual testing or automated testing?
Both. It takes on the manual testing that automation never covered, and records it on the issue. The same sessions export as step-based tests you can replay for regression testing.
Which agents work with it?
Any MCP client: Claude Code, Cursor, Codex, Antigravity, VS Code and others. Settings › Connect an agent in the extension prints the setup for each one.
Which browsers does it support?
Chrome, Microsoft Edge and other Chromium-based browsers, installed from the Chrome Web Store.
Can the agent see my other tabs?
No. Agents work only in the tabs of the “Levr” tab group: tabs they open, or tabs you choose to share. Your other tabs, your history and your bookmarks are not read.
How is this different from a headless browser MCP?
A headless browser MCP launches a separate browser, so the agent starts logged out and its evidence stays on the machine it ran on. Levr Browser uses your own browser, limited to the Levr tab group, and saves screenshots, recordings and network logs to the issue, test or run they belong to.
Can it do exploratory and ad hoc testing?
Yes, and that’s what it’s best at. The manual checks that were never worth scripting (exploring a feature, reproducing a bug, verifying a fix by hand) your agent can now do in your browser, and everything it does is recorded on the issue. When a check should run every release, export the session as a test and replay it.
What does it cost?
Levr is free during the preview, Levr Browser included. No credit card.
Add Levr Browser, then ask your agent to test something.
Needs a Levr account. Free during preview.
// try it on one task
Done means proven.
Connect your agent and give it one real task with gates. See what it says when it can’t just say “done.”
Free during preview · No credit card
# connect Claude Code, Cursor, Codex… to Levr
$ npx @levr-one/cli@latest mcp add
# in Claude Code: add the plugin, then start on a task
> /plugin marketplace add BitModern/levr
> /plugin install levr@levr
> /levr:work ENG-412
# run a task’s gates yourself, or in CI
$ levr gates run ENG-412