The browser you share with your agents.
Use it as your own browser. Let AI agents work beside you, in tabs of their own, with your sign-ins and under your rules — and watch, approve or take over at any moment.
Your browser. Their tabs. Your rules.
Agents are good at web work, but handing one your browser usually means handing it everything. Kin Browser gives it exactly what it needs, where you can see it.
Tabs of their own
Each agent task gets a named space and its own framed tabs. They use your sign-ins, but never see or touch the tabs you have open — and close only what they opened.
Watch, approve, take over
A coral frame and a moving cursor show every step, with the agent's reason beside it. It asks before new sites and risky clicks. One click takes control; another stops it.
Nothing it shouldn't have
Agents act on what they can see, by label — no JavaScript, no selectors, no DevTools. Public web only: never your files, your router or your network, unless you open one address for testing.
From download to your first agent task in five minutes.
Six steps, and the first three take less time than reading this. Pick any step to see it.
Welcome to Kin Browser. Agents work in their own tabs here — you see each one live, and can take over any time.
# or from a terminal
# restart them, then: "use kin-browser to …"
Use kin-browser to compare the 14″ MacBook Pro price on three Malaysian shops, with screenshots.
Cheapest is shop-two.my at RM 8,149 [2]. All three in stock. Screenshots attached.
Kin Browser · an agent wants to press “Submit claim” on portal.example.gov.my
# and on your phone, the console's Approve page
One line to connect. Every ability, as a tool.
Whatever your agent speaks — a terminal, MCP or plain HTTP — the same 139 actions arrive, and the skill teaches it how to use them well.
Hacker News — news.ycombinator.com e3 link "Hacker News" e7 link "new" e41 link "Show HN: …" e42 link "193 comments"
A page arrives as a short list of things the agent can act on. It acts by ref — click e42 — and reads only what changed afterwards. Fewer tokens, no scripts, nothing hidden.
Give it the goal. It does the steps.
Kin Browser now does whole tasks with its own model, remembers every site for every agent, and hands a CAPTCHA to your phone instead of giving up. The rules an agent works under still hold, and every task shows what it cost.
Whole goals, with sources
task "…"plans, browses and answers with citations- Site memory: layout and sign-in learned once, shared by every machine
- Bot checks to a person's phone: tap, type, Done, and the agent goes on
- A task board so agents split work without colliding
- Tokens and US$ per task, with a cap, and what refs saved
- Sees charts and canvases; big goals split across parallel tabs
- Scheduled goals that say what changed; a task that worked becomes a recipe
- Show it a task once; answers checked against their sources; handover between agents
Tests that write and fix themselves
- Describe a flow in words; get a passing kin.yaml test
- API contracts: removed fields and slow endpoints, caught before release
- Load-lite: one journey on 20 browsers at once
- Accessibility fixes as patches or a PR comment
- One comment and a status on every pull request
- Visual diffs triaged; the same journey from every region; replays in bug reports
- An exploratory tester, tests that heal, budgets per route, mocks from contracts
From watching to deciding
- Competitor and price dashboards, shared by link
- LHDN MyInvois, SSM, KWSP and PERKESO packs
- Approvals and morning digests on WhatsApp
- A cited brief on each company before your meetings
- Tender closing dates on your calendar; invoices collected and matched to MyInvois
- Review and social listening; a CRM that stays fresh
- Shopee, Lazada and Zalora price index; grants; supplier checks; the weekly brief
Run it everywhere, your way
- Policy as code: one versioned file for every machine
- The signed audit log streamed to Splunk, Datadog or Elastic
- Data residency, with a report
- Your own update channel: only versions you approved
- SCIM from Okta or Entra ID; your own AWS KMS key
- Team budgets, four-eyes approvals, signed recipes and plugins
- Data loss prevention, compliance recording, air-gapped with Claude on Bedrock
$ kin-browser task "Cheapest 14-inch MacBook Pro in Malaysia today, with the shop"
… 6 steps, last: extract · US$0.118
… 14 steps, last: cite · US$0.401
Task t3fa91c02 — done
The cheapest is RM 7,299 at Shop A (MacBook Pro 14, 16 GB / 512 GB).
Sources:
[1] https://shop-a.my/macbook-pro-14 — “RM 7,299.00”
[2] https://shop-b.my/apple/mbp14 — “RM 7,499.00”
14 steps · 96 s · 182,410 tokens · US$0.412
Saved: ~1,204,880 tokens by reading pages as refs
Built for the work your team already does.
Every ability is a command, an MCP tool and an API call — so the same work runs from a terminal, an AI assistant, a script or a Brainkin teammate.
Rows, not paragraphs
Tables, lists and product cards as JSON, just the fields asked for (name, price, stock), across every page of results, with no repeats.
Questions, with answers
An agent asks you to pick an option or type a one-time code, and gets your answer back. A secret answer is never shown again or recorded.
Wait for hours, then wake up
Kin Browser checks a page every few minutes without the agent, and calls a webhook when “In stock” appears or a price drops below your number.
Do it once, reuse it
Save a task that worked as a recipe with parameters. Any agent runs it by name in one step, and finds each button again after small site changes.
Show where it came from
Every fact can come with its quote, address, time and a picture of exactly where it was on the page.
Many pages at once
Up to eight pages opened and read in parallel, for comparing products, listings or articles in seconds.
A whole form in one step
Fields filled by their labels — text, dropdowns, checkboxes — and never submitted for you. It says what the page thinks is wrong.
Read what pages hand out
Downloaded PDFs, Word, Excel and PowerPoint files, CSVs and pictures, as text and tables.
Language, time zone, phone
See a site as a Malaysian phone user would, or a desktop in Singapore — locale, time zone and device per browser.
Test what you build — by hand, with an agent, or in CI — in the same browser, with the same steps. Chromium, Firefox and WebKit. Open your local app with --allow-local localhost:3000.
# kin-tests/shop.kin.yaml
tests:
- name: search finds a MacBook
steps:
- open: /search
- fill: { Search: MacBook }
- expect: MacBook Pro
- noErrors: true
- a11y: { max: { critical: 0 } }
Checks and test runs
Plain-step test files, expectations that wait, retries, JUnit and an HTML report. --repeat finds flaky steps and says why.
One-click bug reports
Steps to reproduce, screenshot, console errors and environment — filed to GitHub, GitLab or Jira.
Visual regression and every screen
Pixel comparison against baselines, and phone, tablet, laptop and desktop in one call.
Crawl and accessibility
Broken links and error pages across a site; alt text, labels, contrast and headings on every page.
Recordings become tests
Export what an agent did to Playwright in TypeScript, JavaScript or Python, with role-and-name locators.
Signed in, offline, mocked
Saved sign-ins, offline and slow-network modes, mocked API answers, test data and an inbox that picks out the code.
Speed and differences
LCP, CLS and page weight with budgets; staging and production compared, text and pixels.
Built for pipelines
kin-browser ci: a Docker image, GitHub Actions and GitLab CI examples, traces of failures, a comment on the PR.
A shared library
Versioned recipes and site notes, synced through your repo or a cloud machine.
The web work revenue teams repeat every week — research, lists, competitors, campaigns, reviews, tickets — done by agents with sources cited, and a person approving what goes out.
# before a call kin-browser research acme.com.my # tenders with your keywords, to Slack kin-browser watch-new https://tenders.example.gov.my \ --keywords "CRM,software" --webhook https://hooks…
Account research, cited
A one-page brief from a company's own site: offers, hiring, recent news, who they name — every line linked to its page.
Signals, tenders, lead lists
Hear when an account hires or changes prices, catch new tenders, and turn directories into rows straight into your CRM.
Competitor tracking
What changed on their pricing and feature pages since yesterday — added, removed, and which prices moved.
SEO, rankings, campaign QA
A fix-first SEO audit, where you rank in Malaysia or Singapore, and a pre-launch check of UTMs, links, tags, consent and speed.
Proof and reviews
Dated screenshots proving your offer shows where it should; new reviews with replies drafted for you to approve.
Into your CRM
HubSpot, Salesforce, Zoho, Pipedrive and Odoo: contacts, companies, deals, notes, tasks and files — updated or created.
Spreadsheets and portals
One action for every row of a sheet, with approvals; recorded portal tasks on a timing where there is no API.
Reproduce, then attach
See what a customer sees — their plan, device, language — and put the report on the Zendesk, Freshdesk, Jira or Intercom ticket.
Your rules, logged
Blocked sites, robots.txt respected, personal data removed, a polite pace per site, and an audit log of every collection.
New in 1.0. One browser for a team and its agents — on a desktop or a cloud machine — with people signed in, sign-ins agents never see, and approvals from wherever you are.
kin-browser people add [email protected] --role admin kin-browser sso google --clientId … --url https://agent… kin-browser vault add billing.example.com --username ops kin-browser chat slack --approvers U0123,U0456
Sign-ins agents never see
Save a sign-in once. The agent says “sign in” and Kin Browser fills it — with the two-factor code — after your yes.
People, roles, single sign-on
Viewers watch, approvers decide, admins set the rules. Sign in with Google or Microsoft; every decision logged by name.
When this happens, do that
A webhook, a new tender, an email or the clock starts a chain of steps — built without code on the Workflows page.
Google and Microsoft 365
Rows into Sheets or Excel, reports into Drive or OneDrive, calendar events — and emails only from drafts a person approved.
From Slack, Teams or your phone
/kin, @Kin, or a notification on your phone's home screen — no console needed.
Encrypted, kept as long as you say
Secrets encrypted with the machine's own key; recordings, reports and logs deleted on your schedule.
A fleet, in your customers' countries
Spread a thousand-row batch over several machines, and search from Singapore, Jakarta or Tokyo as people there do.
Plugins and a marketplace
New abilities in a few lines of JavaScript, fenced to the sites they declare — or one command from the marketplace.
Usage, updates, your language
A dashboard of the hours saved, one-click updates, a first-run tour, and pages in English, Bahasa Melayu and 中文.
For your whole team, not just your laptop.
Run it on a cloud machine your team watches together — the same Grid, with people signed in and every decision on the record.
A vault, not a password in chat
Agents sign in to billing portals and government sites with the two-factor code, and only ever learn “signed in”.
Approve from anywhere
Allow a site, approve a draft or answer an agent from Slack, Teams, or a notification on your phone.
See the hours saved
A usage dashboard of steps, pages, rows collected and decisions, by day, by agent and by action.
Safe to share — by design, not by promise.
Each of these is enforced by the browser itself, and each has a test that fails if it is ever taken out.
- Their own tabsWhen they finish, they close only what they opened.
- Public websites onlyEvery address is checked, so a name that leads inside your network is refused. Testing your app?
--allow-localopens exactly one address. - No scripts, no codeNo JavaScript, CSS selectors or browser internals. Agents act on what they can see.
- Stop it any timeTake control or stop the agent from its tab. Nothing is queued while you hold it.
- Asks before risky clicksPay, delete, send or submit waits for your yes: once, always on that site, or no.
- Secrets stay secretVault sign-ins are filled for the agent, never shown to it, and encrypted on disk.
- Your rules, loggedBlocked sites, robots.txt, personal data removed, a pace per site, and an audit log of every collection.
- Tasks have limitsSteps, minutes or sites before it must ask you to go on.
- Web pages can't drive itThe agent API refuses any request from a web page, including DNS rebinding.
Get Kin Browser Lite. It's free.
Loading the latest release…
Portable — no installation
Everything you can do, written down.
A step-by-step manual for people and the full reference for agents and developers — searchable, and made from the same guides the app ships with.
Getting started
Install, connect your AI apps in one step, and watch your first agent task.
New in 1.3Goal-driven tasks
Whole goals done for you with sources, what each cost, site memory, bot checks to your phone, the task board.
TeamsFor your whole team
People and roles, single sign-on, the vault, approvals from Slack, Teams and phones.
DevelopersTesting and QA
Checks, test runs, bug reports, visual diffs, smoke tests from a crawl, CI.
New in 1.3Enterprise
Policy as code, SIEM export, data residency, your own update channel, WhatsApp approvals.
ReferenceCommands and MCP tools
Every command, every tool, every option — generated from the release itself.
Asked and answered.
Is it Chrome?
It's a browser of its own, built on open-source Chromium — the project Chrome is built from — with its own window, icon, start page and profile. Sites work as they do in Chrome.
Which agents can use it?
Claude Code, Codex and any agent with a terminal (after kin-browser skill install), Claude Desktop, Cursor and any MCP client (one config entry), and your own programs through a local API. Brainkin teammates use it too.
Does it see my other browsers' passwords or history?
No. It has its own profile, empty until you use it. Sign-ins you save in its vault are encrypted with a key of the machine's own.
Can an agent do anything I can?
No. It can read and click on public websites, in its own tabs. It can't run code, open your files or reach your local network, it asks before risky actions when you turn that on, and you can stop it at any time.
Can I use it to test my own app on localhost?
Yes. Start it with --allow-local localhost:3000 and agents may open exactly that address on your machine — nothing else local. Only you can set this.
Can my team share one?
Yes. Run it on a cloud machine with a console your team signs in to — with roles, single sign-on, a shared vault, and approvals from Slack, Teams or a phone. See docs/teams.md.
Do recordings take up space?
Only if you turn recording on. Each step is a small screenshot, kept on your computer only, under a limit you set (1 GB unless you change it), and deleted on your retention schedule.
Is it free?
Kin Browser Lite is free, for as many machines as you like: the browser, every basic action, recipes, watches, the vault, testing and CI. Pro adds goal-driven tasks, AI testing and business automations; Premium adds company controls (fleets, policy, compliance, your own keys); Ultra adds the built-in agent. Each is licensed per machine. See the editions or contact us.
How do I upgrade?
Install Lite, open Settings in its Grid, and type your licence key (KIN-XXXXX-XXXXX-XXXXX-XXXXX), or run kin-browser licence add KIN-…. Kin Browser checks the key, then fetches the paid features for your version over its own encrypted channel, sealed for your computer. A key works on one computer (Mac, Windows or Linux); to move it, ask us to free it, or free it on your licences page.
Can I try before I buy?
Yes, one step at a time: Lite can try Pro free, and Pro can try Premium free, from Settings (or kin-browser trial start). When a trial ends, your own licence comes back. Ultra is by purchase only.
Can I try new versions early?
With an invitation. Our beta and alpha builds go to invited testers: the invitation is a key you type in Settings, like a licence key. Back to the stable release, in Settings, returns you to stable in one click, with the edition your licence gives.
Can Kin Browser be the agent itself?
Yes, with Ultra. Connect your own ChatGPT plan (through Codex, with a code you type at OpenAI) or a Claude API key, then tell Kin Browser what to do in its Grid's Ask Kin box, or with kin-browser agent "…". It works in its own tabs and keeps your limits. Your sign-in stays on your computer. See the built-in agent.
I have an earlier version. What now?
Update to the newest version: it is Lite unless you have a licence. Earlier versions are no longer offered or supported. Kin Browser shows the update in its Grid, or run kin-browser update.
Give your agent a browser you can trust.
Download Kin Browser, run one command, and watch your first agent task in five minutes.