agent browser alternatives
Vercel's agent-browser is a native Rust command-line tool that gives an AI agent a browser to drive, published under Apache-2.0. The alternatives doing the same job are browser-use, Microsoft's Playwright MCP server, Stagehand and invisible_playwright_mcp, and they differ less in what they automate than in how the agent talks to them and which browser ends up running.
One disclosure before the comparison: this wiki is maintained by the people behind invisible_playwright_mcp, one of the tools below. Read our row with that in mind. The picks at the end name two other tools before they name ours.
The repository describes it as a browser automation CLI for AI agents, and
the shape of the thing follows from that word: it is a compiled binary an
agent runs as a separate process, not a library it imports. A daemon
persists between commands so a session survives from one invocation to the
next, and --json makes the output parseable rather than pretty.
Two details matter when you compare it with the rest. First, an agent can also reach it over the Model Context Protocol, because the binary can run as an MCP server, so it is not CLI-only in practice. Second, it drives local Chrome or a cloud browser provider, which puts it in the same Chromium-family camp as most of this category.
Its feature list leans toward what an agent needs rather than what a test suite needs: semantic locators and element refs, accessibility audits, network interception and multi-session handling. That is a different emphasis from a testing framework, and it is the clearest signal of who the tool is for.
Every cell below was read from the project's own repository on 5 September 2026. Licenses and runtimes change, so re-check before you standardise on one.
| Tool | License | Browser it drives | How an agent talks to it | Runtime |
|---|---|---|---|---|
| Vercel agent-browser | Apache-2.0 | Chrome locally, or a cloud browser provider | CLI with JSON output, or as an MCP server | Rust binary |
| browser-use | MIT | Chromium family, through Playwright | Python library; the agent loop decides each action | Python 3.11+ |
| Playwright MCP | Apache-2.0 | Chromium, Firefox or WebKit | MCP server; any MCP client calls its tools | Node.js 18+ |
| Stagehand | MIT | Chromium, through Playwright | Library with act, observe and extract calls beside raw Playwright | TypeScript, Python or Go |
| invisible_playwright_mcp (ours) | MIT | Firefox, patched at the source level rather than a stock automation build | MCP server, or a local two-pane UI | Python 3.11+ |
Three honest notes on that table. Playwright MCP reads the page's accessibility snapshot instead of screenshots, which keeps each step cheap when a page exposes good semantics and hurts when it does not. Stagehand is the only one that lets you mix plain-language steps and hand-written Playwright calls in the same script, which is the right shape if you already have Playwright code. And browser-use has by some distance the largest community in this list, which matters more than any feature comparison when you hit a problem at midnight.
Process shape. A CLI binary and an MCP server are both out-of-process, so a crash in the browser layer does not take your agent with it. A library is in-process, which is simpler to debug and easier to bring down.
Who decides the next action. browser-use and Stagehand put a model in the loop by design. Playwright MCP and agent-browser expose actions and let whatever assistant you already run do the deciding. invisible_playwright_mcp does both, depending on which of its two entry points you use.
Language. This is the boring criterion that decides most real adoptions. Node-only or Rust-only rules a tool out of a Python shop as effectively as any missing feature.
- You already run an MCP-first assistant and want the plainest hookup: Playwright MCP. It is Microsoft's own, it speaks three browser engines, and there is nothing extra to learn.
-
You have Playwright code and want AI-directed steps inside it:
Stagehand, because
act()sits in the same file as your selectors. - You want the biggest community and a Python agent loop: browser-use. Its own tradeoffs are covered here.
- You want a compiled binary with no runtime to install and a daemon that holds sessions: agent-browser.
- You care that the browser presents as an ordinary desktop browser, and you want a run to repeat identically: invisible_playwright_mcp, ours, conflict noted.
What is Vercel's agent-browser? A native Rust CLI, Apache-2.0 licensed, that gives AI agents browser automation: navigation, semantic locators, screenshots, network interception and multi-session handling, with a daemon that persists between commands.
Is agent-browser open source? Yes, under Apache-2.0 as published in the vercel-labs repository.
What is the difference between agent-browser and browser-use? agent-browser is a compiled binary an agent shells out to or reaches over MCP, with no model inside it. browser-use is a Python library with an LLM-driven agent loop built in: you give it a task, it decides the actions.
Does Playwright have its own MCP server? Yes. Microsoft publishes it under Apache-2.0, it needs Node.js 18 or newer, and it drives Chromium, Firefox or WebKit while exposing accessibility snapshots rather than screenshots.
Is Stagehand the same as Playwright? No. Stagehand builds on Playwright and adds act, observe and extract, so a step can be written in plain language instead of a selector. It ships TypeScript, Python and Go SDKs.
See also: browser-use alternatives for the most-adopted option above, Stagehand vs browser-use for a direct comparison of two rows, and choosing an AI browser agent for the decision framework these picks come from.
- Vercel agent-browser repository, https://github.com/vercel-labs/agent-browser - description, Apache-2.0 license, Rust runtime, CLI and MCP interfaces, daemon and feature list. Read 5 September 2026.
- browser-use repository, https://github.com/browser-use/browser-use - MIT license, Python 3.11 or newer, Playwright underneath, LLM-driven agent loop. Read 5 September 2026.
- Playwright MCP repository, https://github.com/microsoft/playwright-mcp - Apache-2.0, Node.js 18 or newer, accessibility-snapshot approach, browser coverage. Read 5 September 2026.
- Stagehand repository, https://github.com/browserbase/stagehand - MIT license, TypeScript, Python and Go SDKs, act/observe/extract on top of Playwright. Read 5 September 2026.
From the invisible_playwright_mcp wiki. invisible_playwright_mcp is one row in the table above, not the first, and the picks name two other tools before they name ours.