Browser agents

The job

Software that lets a model operate a screen covers more than web pages. Some tools drive a browser, and others tap through a phone or click around a desktop. People building agents look at them when a task has no API.

How they differ

One kind is a library or command-line tool for agents that browse. Browser-use is a Python library, jev-ultrafast reads each page as a table of elements, and webcmd keeps a local memory of sites.

A second kind drives a browser engine directly. Obscura is a Rust headless engine for Puppeteer and Playwright scripts, chrome-devtools-mcp is an MCP server for a running Chrome, and invisible_playwright_mcp wraps a Firefox built to avoid bot detection.

A third kind works inside your own browser. BrowserSkill uses your Chrome or Edge profile and its existing logins, and ego-lite is a desktop browser that can import Chrome logins.

A fourth kind is a full agent app. Atomic-agent is local-first, OpenBot is a self-hosted workspace with a browser for agents, and nanoMuse is one agent on phone, desktop and browser. Chat-on-steroids connects ChatGPT to a machine over MCP.

A fifth kind controls phones and desktops. Mika is an Android app that operates other apps, phone-harness lets a coding agent use a real handset, esp32-iphone-agent taps an iPhone through a board posing as a mouse, and lcu adds desktop control to other harnesses.

Licences

On 10 November 2026, GitHub detected MIT on most of the tools here. Apache-2.0 was on chrome-devtools-mcp, Mika, obscura and webcmd; nanoMuse carried GPL-3.0. A licence can change, and each tool's own page gives the latest reading with its date.

What to check first

BrowserSkill's README says its agent window shares your profile's logins and has no sandbox. Ego-lite offers to migrate Chrome data, and phone-harness gives control of a real phone with apps already signed in. Browser-use warns that domain limits are not a sandbox and that local modes open an unauthenticated debugging port. Jev-ultrafast asks for Chrome remote debugging and its demos call paid model services. Esp32-iphone-agent can operate any app, and its screenshots go to the vision model you configure. Chrome-devtools-mcp warns that its client can inspect and change browser data, and lcu says its cross-turn setting is no security boundary.

Mika acts unprompted except before spending money and sends screen contents to its model provider. NanoMuse passes conversation text through a relay. Browser-use defaults to its hosted gateway and cloud browser.

Invisible_playwright_mcp is built to evade bot detection and has no login if its host is changed. Obscura describes anti-detection features. Mika's README warns that automating other apps may break their terms of service. Chat-on-steroids is an independent beta whose extension automates the ChatGPT web interface, which the project says may conflict with provider terms.

Atomic-agent says secret redaction is not full isolation, and alpha openmuse uses one shared access key. OpenBot says its agents' browser holds real logins and should stay on loopback.