Open-source agents give engineers the parts: browser control, screen grounding models, sandboxes, and agent loops. They are free to run, inspectable, and moving fast. What they leave to you is the product around them: credentials, review screens, rules, scheduling, evidence, and support.
Browser agents and infrastructure
Browser Use and Skyvern
Claims.AI agents that automate browser workflows.
Built for.developers automating web forms and portals.
Reaches software.Browser Use drives Chrome through CDP; Skyvern pairs Playwright with vision models.
Runs.self-hosted or each vendor's cloud. MIT and AGPL-3.0.
Back-office fit.web only. Compared in depth on Browser Agents →.
Stagehand (Browserbase)
Claims.an SDK to extract data and interact with any website, mixing code with natural-language actions.
Built for.developers who want deterministic code with AI fallback.
Reaches software.browser through CDP with Playwright-style methods.
Runs.locally or on Browserbase's cloud browsers. MIT.
Back-office fit.reliable web automation; browser only.
Steel
Claims.an open-source browser API for AI agents and apps.
Built for.teams that need hosted browser sessions.
Reaches software.a Chrome sandbox you control with Puppeteer or Playwright.
Runs.self-hosted with Docker, or Steel's cloud. Apache 2.0.
Back-office fit.infrastructure, not an agent.
Notte
Claims.cloud browser infrastructure and web automation for AI and coding agents.
Built for.developers needing browsers, credential vaults, and scheduled functions.
Reaches software.Playwright locally, CDP sessions in its cloud.
Runs.locally or Notte's cloud. Source-available under SSPL.
Back-office fit.web only.
LaVague
Claims.a large action model framework for building web agents.
Built for.developers.
Reaches software.Selenium, Playwright, or a Chrome extension.
Runs.locally. Apache 2.0; little activity since early 2025.
Back-office fit.web only, quiet project.
Desktop and computer-use agents
Agent S (Simular)
Claims.an open framework that uses computers like a human.
Built for.researchers and developers.
Reaches software.screenshots plus a grounding model that turns intentions into mouse and keyboard input on Windows, macOS, and Linux.
Runs.locally with your model provider. Apache 2.0.
Back-office fit.strong general desktop control; research-grade.
UI-TARS and Agent TARS (ByteDance)
Claims.an open multimodal agent stack for your own computer and browser.
Built for.developers wanting open-weight GUI control.
Reaches software.a vision-language model plus mouse and keyboard; browser by page structure or vision; MCP tools.
Runs.locally. Apache 2.0.
Back-office fit.capable; review data handling before regulated use.
Microsoft UFO
Claims.a desktop agent operating system for Windows, expanding to multi-device orchestration.
Built for.Windows automation research.
Reaches software.Windows UI Automation, Win32 and COM controls, plus vision.
Runs.locally on Windows. MIT.
Back-office fit.well suited to Windows thick clients; needs engineering to harden.
OpenAdapt
Claims.show a task once and it compiles a local program that replays with new data and no model calls at run time, stopping to ask a person on a mismatch.
Built for.desktop apps, EHRs, and payer portals.
Reaches software.accessibility APIs on Windows, macOS, and Linux; Playwright for web; pixels for remote desktops.
Runs.locally. MIT core, with a separate commercial cloud.
Back-office fit.close in spirit to back-office work; a small project with limited published testing.
Cua
Claims.open infrastructure for computer-use agents: drivers, sandboxed desktops, and benchmarks.
Built for.teams building their own desktop agents.
Reaches software.virtual machines and containers for macOS, Linux, and Windows; model-agnostic.
Runs.your machines or your own cloud account. Mostly MIT.
Back-office fit.the sandbox layer, not the task logic.
Self-Operating Computer (OthersideAI)
Claims.a framework that lets multimodal models operate a computer.
Built for.experimentation.
Reaches software.vision plus mouse and keyboard, with OCR modes.
Runs.locally. MIT; low activity since 2025.
Back-office fit.proof of concept.
Open Interpreter
Claims.now a coding agent for open models, rewritten in 2026.
Built for.developers.
Reaches software.code and shell, with skills that can drive browsers and native apps.
Runs.locally. Apache 2.0.
Back-office fit.no longer a computer-control tool at its core.
Side by side
Project
How it reaches software
Where it runs
Legacy desktop or ERP screens
Browser Use, Skyvern
Browser via CDP or Playwright
Self-hosted or cloud
No
Stagehand
Browser via CDP
Local or Browserbase
No
Steel
Browser sandbox
Self-hosted or cloud
No
Notte
Browser via Playwright or CDP
Local or cloud
No
LaVague
Browser
Local
No
Agent S
Screenshots plus grounding model
Local
Yes, research-grade
UI-TARS, Agent TARS
Vision model, browser, MCP
Local
Yes, research-grade
Microsoft UFO
Windows UI Automation and vision
Local Windows
Yes, Windows
OpenAdapt
Accessibility APIs, pixels
Local
Yes, early
Cua
VMs and containers
Your machines or cloud
Sandbox layer
Self-Operating Computer
Vision plus input
Local
Proof of concept
Open Interpreter
Code and shell
Local
No
LaunchAI
The screen of any app, plus APIs
Your machine or an enrolled agent PC
Yes, supported product
When LaunchAI fits better
Pick an open-source project when you have engineers who want to own the loop, are building a product, or are researching computer use.
Pick LaunchAI when the process owner should build the agent by talking and showing, and you want the parts around the model in the box: exact rules, a review step, a schedule, a screenshot of every action, and a human to help inside the tool. See Product →.
