Skip to content
BrowserZap

Private beta · Chrome · any MCP agent

Your AI agent, in the Chrome you already use

Claude Code, Codex or any MCP agent works in your own logged-in tab. It clicks exactly what it means, waits for your approval before anything that submits, sends or pays, and tells you plainly when a site is slow and why.

Get early access See it work

Runs on your computer. No cloud browser, no copied cookies, no telemetry.

BrowserZap in action

shop.example/running/trail-shoes

Running › Trail shoes

Trail shoes

Loading your orders…

Trail Runner 1

$79.00

Add to cart

Trail Runner 2

$89.00

Add to cart

Trail Runner 3

$99.00

Add to cart
Cart · emptyTrail Runner 2 · $89.00 Place order

Your agent

→ page_inspect
3 × button "Add to cart", each named by its card: within "Trail Runner 1", "Trail Runner 2", "Trail Runner 3" · ~1 KiB
→ element_click { name: "Add to cart", within: "Trail Runner 2" }
ok · Clicked button "Add to cart" in "Trail Runner 2"; the page changed. No guess: without within it would have answered TARGET_AMBIGUOUS with the three candidates.
→ element_click { name: "Place order" }
needs_user · USER_CONFIRMATION_REQUIRED: "Place order" is a financial action. Nothing was clicked. Ask the user to approve it in the side panel.
→ element_click { name: "Open order", timeoutMs: 60000 }
ok · waited 28 412 ms · likelyCause: server_slow (still waiting for shop.example/api/orders) · progress sent every 5 s
→ workflow_run { name: "reorder-shoes", paramSets: [ …10 people ] }
8 completed · 2 need attention: #5 stopped at step 3 field_fill "Size" (TARGET_NOT_FOUND, likely not_on_page) → rerun with startAt: 2

An illustration of real BrowserZap results. Click the tabs.

0

wrong clicks in 20 look-alike picks

~1 KiB

per decision, vs 55–106 KiB snapshots

80/80

real-browser tests passing

From our published comparison with simulated agents. Method and limits

Everything an agent needs to work in your browser

Exact targeting, or nothing

Three "Add to cart" buttons? Each is named by its card, row or form. An ambiguous target is refused with the candidates, never guessed.

TARGET_AMBIGUOUS · 3 matches
1 · button "Add to cart" in "Trail Runner 1"
2 · button "Add to cart" in "Trail Runner 2"
3 · button "Add to cart" in "Trail Runner 3"

You approve what matters

Submit, send, delete and pay wait for you, once, for that button on that site.

Click "Send invoice"?

Approve onceDeny

Honest on slow sites

Waits up to two minutes when you allow it, and says why it waited.

  • server_slow · api/orders · 28 s
  • still_rendering · "Loading…"
  • server_error · HTTP 503 · retryable
  • not_on_page · nothing loading

Record once, replay many

Do a task once yourself, or let your agent. Replay it from the side panel with new values, even with no agent connected.

Recording · 4 steps

Open · Search · Fill {{email}} · Save

Batches that don't stop Pro

Fifty people or rows in one run. A failure is reported with its step and cause, and the rest carry on.

✓ #1 · ✓ #2 · ✓ #3 · ✓ #4
✗ #5 step 3 "Email" · not_on_page

Work and personal profiles

Connect every Chrome profile you use. An agent picks the account it acts in, and BrowserZap refuses to guess when more than one is connected.

Work · this sessionPersonalPROFILE_REQUIRED until chosen

Hands CAPTCHAs to you

Never types passwords, card numbers or one-time codes. It pauses, you finish, it carries on.

Works with the agent you already use

BrowserZap speaks the Model Context Protocol. One line of configuration gives your agent 26 task-level tools: no screenshots to parse, no selectors to guess.

Claude Code · testedCodex CLI · testedCursorClaude DesktopWindsurfGemini CLICline
  1. 1Look at the page page_inspect
  2. 2Act exactly element_click field_fill
  3. 3Wait for what matters wait_for_state
  4. 4Save it for next time workflow_record
  5. 5Run it for everyone workflow_run

Local-first, on purpose

There is no BrowserZap server. Nothing you do passes through us, because there's nothing to pass through.

Loopback only
The bridge listens on 127.0.0.1 and needs a pairing token. It refuses to start otherwise.
No code execution tool
Agents get task-level tools. There is no "run this JavaScript", ever.
Nothing you type is logged
Logs and reports record what was clicked and why it failed, never values.
Screenshots stay home
Kept on your computer for 7 days, then deleted. Never uploaded by BrowserZap.

Security · Privacy

Free core. Pro for the repeat work.

Core

Free

Open source, MIT

  • ✓ All 26 agent tools
  • ✓ Approvals, exact targeting, slow-site diagnosis
  • ✓ Record and replay up to 10 workflows
  • ✓ Several Chrome profiles

Pro pilot

Early access

Pricing set with the first users

  • ✓ Unattended batches with reports
  • ✓ Unlimited workflows and version history
  • ✓ Run reports you can review later
  • ✓ Licence checked offline
Join the pilot

Questions

Which agents work with it?

Any client that speaks MCP over stdio. Claude Code and the Codex CLI are tested end to end; configuration snippets are included for Cursor, Claude Desktop, Windsurf, Gemini CLI, Cline and others.

How is this different from the browser control in my AI app?

That works inside one vendor's app. BrowserZap works with whichever agent you choose, and adds what repeated work needs: exact targeting, approvals, workflows you replay from the side panel, and batch reports.

Can it log in or solve CAPTCHAs for me?

No. It hands those steps to you, then carries on. It never types passwords, card numbers or one-time codes.

Does my data go to BrowserZap?

No. Everything runs on your computer. What your agent reads goes to the agent you chose, under its own terms. How data flows

What platforms does it run on?

Google Chrome with Node.js 20 or newer. It is tested on Windows; macOS and Linux builds are on the way.

Show it once. Let it run.

BrowserZap is in a private beta. Tell us which agent you use and what you would automate, and we'll send you a build.

Email [email protected] Read the docs