New v0.22.0 is out

Your agent can finally see the page.

An agent, your shell and a live browser in one window. It edits the code, opens the page and checks its own work.

An animated Conn window. Someone asks the agent to fix a Create account button that is shorter than the email field. The agent reads and edits the signup file and the dev server in the terminal reloads. Conn offers to open the local server, the agent opens the page in the browser tab, clicks the button and takes a screenshot. The button grows from 39 to 44 pixels and lines up with the field.

Bring the agent you already pay for. No API key, no new plan.

  • Claude Code
  • Codex
  • Cursor
  • Grok
  • OpenCode
  • Copilot

How it works

One loop. Change, look, verify.

Most agents stop at "it should work now." Conn's agent sits next to a shell and a live browser, so it runs the app and checks the result before it says it's done.

  1. 01

    Edits the code

    Reads the file, makes the change, and the dev server in the terminal reloads on its own.

  2. 02

    Opens the page

    Navigates the built-in browser, clicks, types and reads the accessibility tree like a user would.

  3. 03

    Checks its work

    Takes a screenshot, measures, reads the console. If it's wrong, it goes around again.

Features

Built for agents that look before they leap.

01 / 03

Point at it. Don't describe it.

Click any element in the preview and it lands in your message with its selector, size and a screenshot.

CtrlShiftE

02 / 03

You decide how much it can do.

Four permission modes, the same for every agent. Try them.

ShiftTab

Asks before each edit, command or page change.

npm test -- signupAllowAlwaysDeny

03 / 03

And the small things that add up.

The parts you stop noticing because they just work.

  • Dev servers open themselves

    A local URL in the terminal becomes a one-click preview.

  • Shared with claude --resume

    Start a chat in the terminal, finish it here.

  • Records the page

    A video of the flow when a screenshot won't do.

  • Eight projects, one window

    Each keeps its own chats, shells and tabs. Pin the ones you use most.

CLI & MCP

Any shell can drive the browser.

Scripts, tests and other agents get the same browser through conn, or through MCP with conn setup project.

  • Accessibility snapshots with stable refs
  • Click, fill, screenshot, record
  • Same pane you're looking at
zsh · ~/shop
$ conn go 3000
loaded localhost:3000/signup
$ conn snapshot
- textbox "Email" [ref=e3]
- button "Create account" [ref=e4]
$ conn fill e3 "you@example.com"
$ conn click e4
navigated to /welcome

Download

Get Conn.

Version 0.22.0 for 64-bit Linux and Windows, plus a phone companion for Android. No macOS build yet.

  • Your system

    Linux

    Debian and Ubuntu

    conn-0.22.0-amd64.deb Download
  • AppImage

    Any other distro

    conn-0.22.0-x86_64.AppImage Download
  • Your system

    Windows

    10 and 11

    conn-0.22.0-x64.exe Download
  • Android

    Phone companion

    conn-mobile-0.1.3.apk Download93 MB

Linux terminal

curl -fsSL https://raw.githubusercontent.com/umarbashirr/Conn/main/install.sh | sh

Windows PowerShell

irm https://raw.githubusercontent.com/umarbashirr/Conn/main/install.ps1 | iex

The script upgrades in place when you run it again. Add --user to install without a password.

FAQ

Questions

Do I need an API key?

No. Conn runs the agent CLI you already have, with its own login. If claude --version works in your terminal, it works in Conn.

How is this different from an editor agent?

Editor agents sit next to your code. Conn's agent sits next to a shell and a live browser, so it can run the app and look at the result before it says it's done.

Is Auto mode a sandbox?

No. It stops on a short list of dangerous commands, which is a convenience, not isolation. For real isolation, run the CLI in an OS-level sandbox.

What does it cost?

Nothing. Conn is open source under the MIT license.

Let it check its own work.