kinbrowserdocs

Getting started

Kin Browser is a browser you use as your own, which AI agents use too — in tabs of their own, with your sign-ins and under your rules. This page takes you from download to your first agent task.

Install

macOS. Open the .pkg and follow the installer. Then open Kin Browser from Applications or Launchpad. The installer also adds the kin-browser command for your terminal. A build that isn't notarised (a developer build) makes macOS ask the first time: Control-click the app, choose Open, then Open again. Releases are signed with a Developer ID and notarised by Apple: see docs/DEVELOPING.md, Signing and notarising.

Windows. Unzip, then double-click Install Kin Browser.cmd. It installs for you only, with no administrator rights. Kin Browser appears in the Start menu and on the desktop, and in Settings → Apps, where you can uninstall it. The installer also adds the kin-browser command for new terminals. Until the app is signed, Windows SmartScreen may ask first: choose More info → Run anyway.

Linux. sudo apt install ./KinBrowser-linux-x64.deb (installs to /opt/kin-browser, with the libraries Chromium needs). Or unpack KinBrowser-linux-x64.tar.gz and run ./install.sh, for you only in ~/.local, or sudo ./install.sh --system for everyone. Kin Browser appears in your applications menu.

Your first agent task, step by step

1. Open Kin Browser. The first time, a short tour explains how agents work beside you and asks whether to send crash reports (off unless you tick it). Pick English, Bahasa Melayu or 中文 from the menu in the bar.

2. Connect your AI apps, once. The first-run tour has a Connect my AI apps button; or, in a terminal:

kin-browser agents connect

It finds the AI apps on this computer and registers Kin Browser in each, so you never edit a config file:

AppHow it learns Kin Browser
Claude Codethe skill, in ~/.claude/skills/kin-browser
Codexan MCP server in ~/.codex/config.toml
Claude Desktopan MCP server in claude_desktop_config.json
Cursoran MCP server in ~/.cursor/mcp.json
VS Code (Copilot agent mode)an MCP server in the user mcp.json
Windsurf, Gemini CLIan MCP server in their settings
Other agentsthe skill, in ~/.agents/skills/kin-browser

Each app's own file is merged, never replaced, and a copy of it as it was is kept beside it (.kin-backup). Restart the apps afterwards. kin-browser agents shows which are connected; kin-browser agents disconnect takes Kin Browser out again. Installed an AI app later? Run agents connect again.

Kin Browser also puts the skill in ~/.agents/skills (and Claude's and Codex's skill folders) by itself each time it starts. What it installs is a short entry that tells the agent to run kin-browser skill print, which prints the guide for the Kin Browser on this computer, so the agent's instructions always match. A skill folder you wrote yourself is left alone. Plugins for Claude Code, GitHub Copilot, Codex, Cursor, Grok Build and OpenCode are at https://kinbrowser.com/plugins/.

The skill tells an agent everything: how to read a page, act by ref, name its space, ask you things, and which commands exist for research, testing and revenue work. For an app not listed, add Kin Browser as an MCP server by hand:

{
  "mcpServers": {
    "kin-browser": { "command": "/usr/local/bin/kin-browser", "args": ["mcp"] }
  }
}

3. Ask for something. In your agent, in plain words:

Use kin-browser to compare the price of the 14-inch MacBook Pro on the three biggest Malaysian electronics sites, with a link and a screenshot for each.

The agent starts Kin Browser if it isn't open (kin-browser start --background), names a space for the task, and opens its own tabs. They have a coral frame and a moving cursor, so you can see each step, and its reason beside it. Your own tabs are never touched.

4. Stay in charge. Turn on Ask before new websites and Ask before risky actions in the Grid, and the agent waits for your yes before a new site, a payment, a send or a delete. When it needs you (a sign-in, a captcha, a choice, a one-time code) its tab says so and you get a notification. Press Take control at any time to pause it, or Stop agent to end it.

5. Let it sign in without your password. Save a sign-in once in the Grid → Vault (or kin-browser vault add billing.example.com --username [email protected]). When the agent reaches that login page it says sign-in, you allow it, and Kin Browser fills the form and the two-factor code. The agent only learns that it worked.

6. Make it routine. When a task went well:

  • Record and save it as a recipe, so any agent can run it by name in one step next time (Recordings → Save as a recipe agents can call).
  • Watch a page or schedule it: "tell me when this tender list has something new", "check this price every hour" (Schedules).
  • Build a workflow: when a webhook is called, a watch fires, an email arrives or on a schedule, run steps one after another — new tender → research the buyer → create the CRM deal → draft an email for me to approve (Grid → Workflows, or ask the agent to automate it).
  • See what it saved you on the Usage page.

7. Bring the team. On a shared or cloud machine, add people with roles and single sign-on, approve from Slack, Teams or your phone, connect Google Workspace or Microsoft 365, and spread big batches over several machines. See docs/teams.md.

Your own program can do all of this too: a local API on 127.0.0.1, protected by a token only your account can read. The verbs are in contract/contract.ts.