Goal-driven tasks, site memory, bot checks and the board (1.3)
Kin Browser 1.3 can do a whole goal by itself, remembers what agents learned about each site, hands bot checks to a person's phone, and lets several agents split one job without colliding. Everything here goes through the same verbs and rules as any agent's work: allowed sites, risky clicks, budgets, robots.txt and the signed audit log all still apply.
Goal-driven tasks
kin-browser ai key # once: your Claude API key, kept encrypted on this machine
kin-browser task "Find the cheapest 14-inch MacBook Pro sold in Malaysia today, with the shop and price"Kin Browser's own model (Claude, claude-opus-5-5 by default) plans the
steps and takes them with the browser verbs (open, extract, cite and the rest),
using your saved recipes and what agents learned about each site. It
comes back with an answer and the sources it rests on:
Task t3fa91c02 — done
Goal: Find the cheapest 14-inch MacBook Pro sold in Malaysia today, with the shop and price
The cheapest is RM 7,299 at Shop A (MacBook Pro 14 M5, 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$0.412 (claude-opus-5-5)
Saved: ~1,204,880 tokens by reading pages as refsThe task runs in the background, and the command waits for it and shows
progress. Add --background to get the id back at once. Then use
kin-browser tasks <id> to see it, tasks <id> --wait 100 to wait for it, or
tasks <id> --cancel to stop it. Over MCP, task takes wait (up to
110 seconds), and you poll tasks with the id.
| Option | Default | |
|---|---|---|
--url | — | Where to start |
--space | — | The space its tabs go in, so you can watch it in the Grid |
--max-steps | 60 | Browser steps |
--max-minutes | 20 | Wall-clock time |
--max-tokens | 1,500,000 | Tokens across all of Claude's turns |
--max-usd | 3 | Dollars, priced per model |
--model | claude-opus-5-5 | Or set it once: kin-browser ai model claude-sonnet-5-5 |
The task stops at whichever limit it reaches first and says which. Your own
limits for every task (the Grid's limits) still count each step too. The
defaults can be set per machine (kin-browser ai limits --max-usd 2) or for
the whole team in the policy file.
A task never types a password: it signs in with the vault (sign-in). It
asks a person for anything that is theirs (a CAPTCHA, a code, a payment, a
choice). It never sends, posts, pays or deletes unless the goal says so.
What each task cost, and what it saved
kin-browser tasks lists every task with its steps, minutes, tokens and US$,
and totals them. The Work page in the Grid shows the same, with each
answer. "Saved" compares two things:
- Tokens: each page as Kin Browser reads it (refs and names) against the same page as HTML. That is what a model would have read without Kin Browser.
- Steps and time: steps a recipe replayed without asking the model.
Memory of sites
Every time an agent opens a site, Kin Browser keeps its main links, its search
box and its sign-in link. When a vault sign-in works there, it keeps where the
login page is. Notes agents and people leave (note) are the site's quirks.
Recipes that start on the site are listed with it.
kin-browser site-memory https://www.lazada.com.myA goal-driven task reads this before it starts, and an agent can too
(site-memory, MCP site_memory). Names only: never what a page says about a
person.
Shared across machines. Point each machine at the team's cloud machine, and site memory travels with the team library (recipes and notes) every 15 minutes. A site's newer facts win:
kin-browser memory sync https://agent.example.com --key kbc_…
kin-browser memory statusCAPTCHA and bot-wall handoff
When a page an agent opens is a bot check, open says so: Cloudflare's "Just
a moment", Turnstile, reCAPTCHA, hCaptcha, DataDome, a "press and hold", or a
plain "access denied". It adds a line at the end:
⚠ A bot check is on this page (recaptcha): A reCAPTCHA: a person solves it.
Hand it to a person: kin-browser handoff --wait 300 (MCP: handoff)handoff hands that one step to a person. Their phone is told (push, Slack,
Teams or WhatsApp, whichever is set up), and the Bot checks section of the
Work page (in the Grid, or on the console from a phone) shows the agent's tab
live:
- they tap on the picture where they would click, type into the box, and use Enter, Tab, ⌫ and scroll;
- then they press Done and the agent carries on.
Their taps reach the page as a person's input. They work while the agent
waits, and only from a person's pages (the viewer key, or an approver signed
in to the console), never from the agent API. handoff --wait 300 waits for
them and says whether the check is gone. A goal-driven task does all of this
by itself when it meets a check.
The task board
Several agents (or several machines, through a cloud machine's API) split one job:
kin-browser board add "Page 1 of the tender list
Page 2 of the tender list
Page 3 of the tender list"
kin-browser board claim # the oldest open card, locked to you for 15 minutes
kin-browser board renew c1a2b3 # still working on it
kin-browser board done c1a2b3 --result "12 tenders, 3 for IT"
kin-browser board # who holds whatA claim is a lock: two agents claiming at once get two different cards. A
claim whose lease runs out is free again, so an agent that crashed does not
hold its card for ever. A card can wait for others (after). Agents that
share one space tell themselves apart with --as "Agent 2". People see the
board on the Work page and can add or remove cards.
New in 1.4
Seeing the page
When the elements do not show what a task needs (a chart, a canvas, a map,
text in an image), the task takes a screenshot and looks at the picture
itself. It can click a point on it (clickAt) or read the text in it
(look).
Big goals, in parallel
kin-browser task "Compare the price of the iPhone 17 at the five biggest Malaysian shops" --parallel 4With --parallel, Claude first splits the goal into independent parts. Each
part becomes a card on the task board. Up to 6 tabs work
through the cards at once, each part as its own task with its share of the
limits. When they are done, the answers are merged into one, keeping every
source. Agents on other machines can claim the same cards through the cloud
machine's board. tasks <id> shows each part's status.
Keep a task as a recipe
A task that worked shows canSave. Keep it:
kin-browser save-task t3fa91c02 laptop-price
kin-browser recipe laptop-price search="MacBook Pro 14"The steps that opened, clicked, typed and read are kept, by element name.
What was typed becomes a parameter, so the next run is a replay: no model,
no cost, and seconds instead of minutes. A recipe that later breaks can be
repaired (repair-recipe) like any other.
Scheduled goals
kin-browser schedule-goal "New IT tenders on ePerolehan this week" --weekday mon --at 09:00 --webhook https://hooks.slack.com/… --whatsapp
kin-browser schedule-goal "The top 5 posts about Acme on Lowyat forum" --every-minutes 360
kin-browser goals # each goal, its timing and its last run
kin-browser goals run g1a2b3 # now
kin-browser goals pause g1a2b3Each run is a task, in the machine's own headless browser, while Kin Browser
runs. Its answer is compared with the last run's, line by line, and what
changed is sent to the webhook and on WhatsApp (+ new, − gone). A run
with nothing new sends nothing on WhatsApp.
Team budgets
Each task counts against a team's monthly budget: --team sales, or the
team its API client is assigned to. When a team's month is spent, its next
task is refused, and a task that starts is capped at what is left. See
Budgets and chargeback.
New in 1.5
Show it once: learning from a demonstration
kin-browser teach start https://portal.example.my --name check-invoice
# … do the task yourself in the tab that opens …
kin-browser teach stopWhile you teach, Kin Browser's layer over that one tab notes what you click
and what you type into which field, by role and name. It never keeps a
password, a one-time code or a card field's value. teach stop turns it into
a recipe: what you typed becomes parameters, and the steps are listed in plain
words for you to check. Run it with
kin-browser recipe check-invoice invoice_no=…. Agents cannot act in the
teaching tab while you teach.
Answers that check themselves
When a task finishes with sources, a second, faster model (Claude Haiku 5.5)
reads the pages it cites and checks each claim in the answer against them.
Each quote is also looked for on its page. The result is shown under the
answer: claims supported, claims not supported, and quotes that were not
found. --verify false (MCP verify: false) skips it.
Watchers with judgment
kin-browser smart-watch add --name "Laptop shop" --urls "https://shop.my/laptops,https://shop.my/deals" --about "price changes on MacBooks, new models, stock running out" --whatsappEvery N minutes (360 by default) the pages are read and compared with last
time. Claude decides whether any change matters to what you said (about).
Dates, counters, ads and rewording are ignored. Only what matters is sent, as
one or two sentences.
A clean browser for a task
kin-browser task "Download this month's statement" --isolated --state finance-admin--isolated runs the task in a separate headless browser that is thrown away
after. With --state, it starts signed in from a saved sign-in
(save-state). Nothing it does stays in your browser, and nothing of yours
leaks into it.
Handing work to another agent
kin-browser handover pack --goal "Compare bank FD rates" --note "Maybank and CIMB done; RHB next"
kin-browser handover take h3fa91c02 # another agent, here or through the cloud machineThe pack keeps the agent's pages (and their spaces), its sign-ins and where it got to, as a card on the task board. Whoever takes it gets the same tabs back, signed in, and is the only one who can take it.