Best Browser MCP Servers for AI Agents

Start with Playwright MCP for general browser control: it is maintained by Microsoft, Apache-2.0 licensed, and drives pages through accessibility snapshots instead of screenshots. Add Chrome DevTools MCP for performance traces and network debugging. Use Browserbase when you want a hosted cloud browser. This roundup compares four servers by maintainer, transport, auth and install line. Checked 31 October 2026.

The short version

Pair pages for single servers have little search demand, so this is a category roundup. The site's listings for each server are linked in the sections below, and the wider directory is at MCP servers.

What a browser MCP server does

A browser MCP server gives an agent tools to open pages, click, type, read content and take screenshots. The agent decides which tool to call, and the server returns the result.

Two designs exist. One reads the page as a tree of roles and labels, called an accessibility snapshot, and acts on elements by reference. The other drives Chrome directly through its debugging protocol and exposes developer tools such as traces and network logs. A third style runs the browser in the cloud so your machine does not.

You choose by what you need to see. For testing a signup flow, the first is enough. For finding why a page is slow, you need the second.

The comparison table

Repository facts on 31 October 2026:

The first two are active, the third is archived, and the fourth has not been pushed to in about ten months.

Playwright MCP

Maintained by Microsoft. The README describes a server that provides browser automation using Playwright, letting an agent interact with pages through structured accessibility snapshots. It lists the main design points: fast and lightweight, no vision model needed, and deterministic tool application.

Setup. It needs Node.js 18 or newer. The README gives install lines for many clients, including claude mcp add playwright npx @playwright/mcp@latest, codex mcp add playwright npx "@playwright/mcp@latest", amp mcp add playwright -- npx @playwright/mcp@latest, VS Code, Cursor and Cline.

Options worth knowing. --browser picks Chrome, Firefox, WebKit or Edge. --headless runs without a window (headed is the default). --isolated keeps the profile in memory. --storage-state loads saved logins into isolated sessions. --user-data-dir sets the profile folder. --allowed-origins lists trusted origins, though the README warns this does not serve as a security boundary. --secrets loads secrets from a dotenv file. --caps adds vision, PDF or devtools capabilities. --port listens for SSE transport. --extension connects to a running Chrome or Edge, which needs the Playwright Extension.

By default it runs a persistent profile like a regular browser, so logins carry over between runs. That is convenient and also a thing to remember when an agent browses on your behalf.

A caveat from the README itself. It points coding agents at a CLI plus skills alternative, and says agents increasingly favour CLI workflows because they avoid loading large tool schemas and verbose accessibility trees. If you are token-conscious, test both. Our listing is Playwright MCP.

Chrome DevTools MCP

Maintained by the Chrome DevTools team. The README says it lets a coding agent control and inspect a live Chrome browser for automation, debugging and performance analysis. It uses Puppeteer for automation and Chrome DevTools to record traces and extract performance insights, and exposes network requests, screenshots and console messages.

Setup. It needs Node.js LTS, npm and current stable Chrome. The server starts the browser automatically when a client first uses a tool that needs one. It officially supports Google Chrome and Chrome for Testing only.

Options. --headless, --isolated and --slim exist; the slim flag exposes a reduced tool set for basic browser tasks, which cuts the tools loaded into context.

Privacy. The README says usage statistics collection, such as tool invocation success rates, latency and environment information, is enabled by default, and you can opt out with --no-usage-statistics. It adds that Google handles this data under the Google Privacy Policy and that the choice is independent of Chrome's own metrics. If you work on private pages, set the flag. Our listing is Chrome DevTools MCP.

Browserbase

Browserbase sells cloud browsers. Its MCP server lets an agent create, reuse and close browser sessions with explicit tools, handle concurrent sessions, and view session replays. The repository for the self-hostable version is archived and says it is no longer maintained, with the hosted server recommended. The hosted server uses Streamable HTTP at https://mcp.browserbase.com/mcp, and the README says Browserbase provides the LLM costs for Gemini on that server. You set Browserbase API credentials in your MCP configuration.

Cloud browsers help when you do not want an agent clicking around on your own machine, when you need many parallel sessions, or when sites block local automation patterns. The cost is a vendor account and usage billing. Browserbase's pricing page lists Free at $0 with 3 concurrent browsers and 1 browser hour, Developer at $20 a month with 25 browsers and 100 hours, and Startup at $99 a month with 100 browsers and 500 hours, with overage at $0.12 and $0.10 per browser hour on the paid plans. Our listing is Browserbase MCP.

ExecuteAutomation Playwright MCP

A community server, MIT licensed, built on Playwright. Its docs include a device quick reference, a prompt guide and a supported-tools list. It installs through claude mcp add --transport stdio playwright npx @executeautomation/playwright-mcp-server and other routes such as Smithery. It is useful if you want its extra tooling, such as API testing helpers. The last push we saw was 13 December 2025, so check the issue tracker before you rely on it for current Playwright versions. The repo is not archived, shows no tagged release and had 46 open issues on 31 October 2026.

Puppeteer and older servers

Older lists still show a Puppeteer MCP server from the official reference set. That reference collection has moved: the modelcontextprotocol/servers README now lists a short set of reference implementations (Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time) and notes an Archived section, and an archived repository holds the older ones. Treat any Puppeteer server as legacy. Our entry, Puppeteer MCP, covers it for completeness. For discovery, the README points to the official MCP Registry at registry.modelcontextprotocol.io.

Which to pick for which job

For pure content extraction rather than interaction, a fetch or crawl server is lighter than a full browser. See Fetch MCP and Firecrawl MCP.

A first task to try

Pick something small and checkable so you can judge the result. Ask the agent to open your staging site, sign in with a test account, add an item to a cart and report the total. Run it three ways.

1. Playwright MCP, default settings. Watch whether the agent finds elements by their labels and whether it needs a second try. Note the headed window: you can watch it work.
2. Playwright MCP with `--isolated`. Confirm that a fresh profile means a fresh login, and that nothing from your real browser leaks in.
3. Chrome DevTools MCP. Ask the same agent to load the page and tell you which request took longest. This is the question the first server cannot answer well.

Record how many tool calls each run used and whether the final answer was correct. That gives you a feel for token cost and reliability that no table can.

Token cost and tool count

Every MCP server adds tool definitions to your agent's context, and a browser server also returns page content. Two things keep that in check. First, load only the servers you need for the task. Second, use the options that shrink the surface: Chrome DevTools MCP has a --slim flag with a reduced tool set, and Playwright MCP lets you enable extra capabilities such as vision, PDF and devtools only when you want them with --caps.

Claude Code defers MCP tool definitions by default, loading only tool names until a tool is used, per its docs, which we cover in MCP vs API. Other agents may load everything up front. If your agent slows down or costs more after you add a browser server, that is the first thing to check. Snapshots of large pages are the other cost: a long page produces a long tree, so ask the agent to work on one section at a time.

Security and privacy

A browser is the most powerful tool you can give an agent. It can read pages you are logged into and submit forms.

• Use isolated profiles for anything you do not trust. Playwright MCP's --isolated keeps the profile in memory.
• Do not point an agent at your main profile with saved logins unless you mean to.
• Page content is untrusted input. A web page can contain instructions aimed at your agent. The MCP specification says tools represent arbitrary code execution, and hosts should get user consent before invoking them.
• Allow lists are not walls. Playwright's own README says its allowed-origins option is not a security boundary.
• Check telemetry defaults. Chrome DevTools MCP collects usage statistics unless you opt out.
• Cloud browsers add a party. Pages and credentials you use there pass through a vendor.

Install for each agent

The Playwright server lists many agents. A few lines from its README:

claude mcp add playwright npx @playwright/mcp@latest
codex mcp add playwright npx "@playwright/mcp@latest"
amp mcp add playwright -- npx @playwright/mcp@latest

For Cursor, VS Code, Cline and others, add the standard config block from the README to the agent's MCP settings. Our Claude Code MCP servers page and best MCP servers list cover agent-specific setup.

Reading the repository facts as risk

The star counts and push dates in the table are a maintenance signal, so use them to plan.

A server that is archived or quiet is not broken. It means nobody is fixing the next incompatibility for you, so assign an owner on your side before you depend on it.

A team rollout in four steps

1. Pick one server for one job. Use Playwright MCP for flow testing and Chrome DevTools MCP for performance and network questions. Two servers with overlapping tools make the agent choose badly.
2. Run it against a staging site with a throwaway account. Never point a first run at production or at an account with real payment details.
3. Write the allowed domains into your instructions file. Tell the agent which hosts it may open. A page can contain text that tries to steer an agent, so a short allowed list is a real control.
4. Pin the version and write down the install line, in the same document where you record your Node version.

When a browser task goes wrong

• The agent clicks the wrong element. Ask it to take a fresh snapshot before each action. Stale references are the usual cause when a page re-renders.
• A page needs a login. Use a dedicated test account and keep its credentials out of the chat. Local servers use the browser profile you configure, so check what that profile can already see.
• Output is huge and the context fills up. Ask for a targeted check, such as one form or one route, instead of a full-page read. The token cost section above explains why tool count and page size both matter.
• The server cannot start. Check the Node version first, since each server is launched through npx. Then run the command by hand in a terminal to read the error.
• Pages load differently for the agent than for you. The browser it controls may have no extensions, a different window size and a clean profile. That is useful for testing and surprising for debugging.

Questions people ask

Do I need a vision model? The Playwright MCP README lists no vision model needed among its design points, because it works from accessibility snapshots.

Can two of these run together? Yes, but give each a distinct job. Overlapping tools make the agent's choice less predictable.

Is a hosted browser better than a local one? A hosted service removes local setup and needs an API key and a trust decision about where your pages are rendered. A local server keeps pages on your machine.

Our take

Start with Playwright MCP. It is the best documented, has the widest agent coverage, and its accessibility-snapshot design works with any text model. Add Chrome DevTools MCP only when you debug performance or the network, and set its opt-out flag if you work on private pages. Choose Browserbase when you want the browser out of your machine, understanding it is a paid hosted service.

We did not run these servers against a shared task, so we make no speed or success-rate claims. The table comes from each repository and docs on 31 October 2026. Two cautions. Token cost is real: tool schemas and page snapshots use context, and Playwright's own README suggests a CLI route for coding agents. And always test with an isolated profile first.

Related pages

• MCP servers directory
• Best MCP servers
• MCP vs API
• Best docs MCP servers