kinbrowserdocs

The built-in agent (Ultra)

Kin Browser can be the agent itself. Type what you want in the Grid's Ask Kin box, or run kin-browser agent "…", and Kin Browser does it by itself. It works in its own tabs, in a space called Kin, on a model account you connect. You don't need Claude Code, Cursor or any other agent.

It is part of Kin Browser Ultra: Premium, and Kin Browser as the agent itself.

kin-browser models connect chatgpt     # your ChatGPT plan, through Codex
kin-browser agent "Compare the 14-inch MacBook Pro on three Malaysian shops, with sources"

Connecting a model account

In the Grid, open Agent settings → Model accounts.

ChatGPT / Codex

This uses your own ChatGPT plan (Plus, Pro, Business or Enterprise), the same way the Codex CLI does. There is no API key and no extra bill.

  1. Press Connect ChatGPT (or run kin-browser models connect chatgpt).
  2. Kin Browser shows a code. Open auth.openai.com/codex/device, sign in to ChatGPT and type the code.
  3. Kin Browser asks your account which models it serves, and uses one of them. Choose another one in the list if you like.

The sign-in stays on this computer, encrypted, and is never sent to kinbrowser.com. Kin Browser renews it by itself. If OpenAI stops accepting it, the account shows Needs connecting again, and tasks stop until you connect it again.

Claude

Paste an API key from console.anthropic.com (or run kin-browser models connect claude). Anthropic bills you for its use. Companies can run Claude in their own Amazon Bedrock or Google Vertex AI account instead, through their team policy (see enterprise.md).

There is no "Sign in with Claude": Anthropic's terms do not allow other apps to use a Claude.ai subscription. To work on your Claude plan, use Claude Code itself: Settings → Use with Claude Code → Connect gives your Claude Code Kin Browser's skill and tools, and it drives this browser on your plan.

Which one runs

With both connected, Use this one chooses. kin-browser models shows what is connected and which one is used.

What it does, and the rules it keeps

An instruction becomes a goal-driven task (see tasks.md). The agent plans, opens pages, reads, clicks and fills forms, through the same actions any agent uses. So everything that holds for an agent holds for it too:

  • the sites you allowed, and your answers to risky clicks;
  • the task limits (steps, minutes, tokens);
  • robots.txt, the audit log and the recording;
  • for a company, the team policy.

Follow a task on the Work page, or with kin-browser tasks.

  • No silent switching. If your connected account fails (its plan limit, a refused model, a sign-in OpenAI no longer accepts), the task stops and says why. It never moves to another account or key by itself.
  • A model the account refuses is remembered and not asked for again. Run the task again to use another one.
  • Every call is counted against the account, in tokens, even though your subscription pays for it. On a ChatGPT account the cost shows as US$0.

More than one question (1.14)

  • About this tab: open the side panel (Kin Browser's toolbar button) and use Ask Kin about this tab. You share the tab in front of you for that one task; Kin works in it, then gives it back, never closing it.
  • A conversation: after an answer, your next message follows up on it, on the same tab, with the questions and answers so far. New conversation starts afresh.
  • Routines: say how often — "every weekday at 8, …", "every hour, …" — and Ask Kin sets a routine instead of doing it once. Ask for "a PDF each time" to keep each report in your save folder, and "tell me only if it changed" for a notification only then. See them under Tasks & questions → Routines.
  • Files: ask for a spreadsheet, a Word document or slides, and they are saved in your save folder (Downloads unless you chose another). Ask it to "handle every PDF I put in ~/Invoices" and that folder is watched: each new file is a task. See local-files.md.
  • From your phone: Settings → Phone & cloud connects a Telegram bot of your own; message it, and the answer comes back with the report as a PDF.
  • On your cloud machine: add it in the same place, then tick Run on my cloud machine under Ask Kin.

Making it yours (1.14)

  • Starters: under Ask Kin, pick your role for ready instructions; fill in the [bracketed] parts.
  • Show Kin how: do a task once in a tab and press Done; Kin keeps it as a recipe to run now or as a routine.
  • Memory: Settings → Memory holds your profile, your company's and what to remember. Ask Kin uses them, and fills forms from them; you can change or delete anything there.
  • Models: a quicker model for simple steps (ChatGPT or local), Claude carrying on when ChatGPT stops (if you tick it), or a local model (Ollama, LM Studio) so nothing goes to a model company. Settings → Model accounts.
  • Approvals from your phone: with Telegram connected, what waits for you arrives with Allow and Decline buttons.
  • What it saved: Tasks & questions shows the time and cost, by routine too.

Around your day (1.14)

  • Speak: the 🎤 button records what you say into the box; voice notes to your Telegram bot work too. Settings → Voice chooses who transcribes.
  • Inbox and calendar: on your Gmail, Outlook or calendar tab, the side panel offers the usual asks. Replies and invitations are drafts you approve; nothing is sent without your yes.
  • Results where you work: "add these to my Notion database", "post it to the zapier webhook" — to what you connected in Settings → Connections.
  • Before it runs: Ask Kin shows what a task will likely take, from your own history; each routine shows its cost a month.
  • Routine rules: a monthly budget, quiet hours, "ask me before it submits", and a pause after failures.
  • Your team: Settings → Share makes a Kin pack of recipes and routines to send; a teammate brings it in (routines paused).
  • Your phone: the phone app, or your cloud machine's /pages/phone.html.
  • Trust: each answer says how far it was checked against its sources.

Commands

kin-browser modelsWhat is connected, and which account the agent uses
kin-browser models connect chatgptConnect your ChatGPT plan with a code you type at OpenAI
kin-browser models connect claudeConnect your Claude API key (asked for; it is not shown as you type)
kin-browser models use chatgpt / claudeChoose which account runs
kin-browser models model <name>The ChatGPT model to ask
kin-browser models refreshAsk the ChatGPT account which models it serves
kin-browser models disconnect chatgpt / claudeTake it off this computer
kin-browser agent "…"Give Kin Browser an instruction (the same as kin-browser task "…")