Code, design and browser MCP servers, compared

Filed under browser automation everywhere, and automation is the least of it. Traces, Lighthouse, heap snapshots and source-mapped console output are the part nothing else on this list has.

What it attaches to

Its own Chrome by default, launched against a dedicated profile under $HOME/.cache/chrome-devtools-mcp. Optionally your running browser instead, via --autoConnect (Chrome 144+) or --browserUrl against an open debugging port.

What it inherits by attaching

Nothing, until you attach it to your own browser — and then every open window in that profile, signed in as you. Opening the remote debugging port is the bigger step: it is unauthenticated, so any process on the machine can drive that browser while it is open.

The lean choice for driving a page. It works off the accessibility tree rather than screenshots, so no vision model and deterministic structured output.

What it attaches to

A Playwright-managed browser it starts itself — headed by default, --headless or PLAYWRIGHT_MCP_HEADLESS to change that. Every CLI flag has a matching PLAYWRIGHT_MCP_* variable, which is the cleaner way to configure it in a container.

What it inherits by attaching

Deliberately little. File access is fenced to workspace roots and file:// navigation is blocked, until --allow-unrestricted-file-access removes both. Microsoft's own caveat is worth reading first: for a coding agent, the Playwright CLI with skills may beat MCP, because it keeps large tool schemas and accessibility trees out of the context window.

The other server called Playwright MCP. It records a session into a runnable test file and emulates 143 named devices — and it has not been pushed since December 2025, which nothing at install time tells you.

What it attaches to

A Chromium, Firefox or WebKit browser it launches itself, downloading the binaries on first use. Headed by default: headless is false unless you set it. There is also a standalone HTTP mode on port 8931 for display-less machines, which needs "type": "http" in the client config or it fails with a 400.

What it inherits by attaching

A visible local browser plus playwright_evaluate for arbitrary JavaScript — and, unusually, playwright_get/post/put/patch/delete, which fire HTTP requests with no browser involved. The HTTP server binds to localhost only by design; remote access is an SSH tunnel.

The clearest example in this cluster of a server that is useless without something already running — and the reason most "it connects but nothing happens" reports exist.

What it attaches to

A socket on port 9876 opened by an addon running inside Blender. Blender has to be open, the addon enabled, and "Connect to Claude" pressed in the sidebar before any tool call can do anything.

What it inherits by attaching

Your live scene, and through execute_blender_code, arbitrary Python inside Blender. That is as powerful and as dangerous as it sounds — the project says to use it with caution outside experiments and to save your work first.

The npm package everybody pastes does not exist. @github/mcp-server returns 404 on the registry, so most circulating instructions fail on the first run.

What it attaches to

A hosted remote server at api.githubcopilot.com with nothing to install, or a Docker image, or a Go binary you build. Take the remote one unless you are on GitHub Enterprise Server, which it does not cover.

What it inherits by attaching

Exactly what your token or OAuth grant carries — this is the one server here where the read-only guard is a real, documented switch. Connect to /mcp/readonly on the remote server, pass --read-only to the binary, or set GITHUB_READ_ONLY=1 on Docker.

Two servers with one name, and unusually the community one is the larger. GitLab ships an MCP server inside the application; zereight/gitlab-mcp ships 217 tools against the API. Which you can use is decided by your instance version, not by preference.

What it attaches to

The built-in server attaches to your GitLab instance itself, at https://<instance>/api/v4/mcp, with no package installed — HTTP transport direct, or npx mcp-remote for clients that only speak stdio. The community server attaches to the REST and GraphQL APIs from a local npm process.

What it inherits by attaching

The built-in server inherits an OAuth grant it registers for itself through Dynamic Client Registration, so nothing long-lived sits in a config file — but it needs three instance settings on (Duo availability, beta features, MCP access) before it answers at all. The community server inherits a personal access token in plain text and defaults to full access, including deletes through execute_graphql; GITLAB_PERMISSION_MODE=readonly or modify is the switch, and there is no equivalent on the built-in one.

Most listings still hand you the desktop setup, and Figma's own docs no longer lead with it. About a third of the tools exist only on the remote server.

What it attaches to

The remote server at mcp.figma.com/mcp, which is now the recommended one, or the desktop server running inside the Figma app. The desktop server's one advantage is that it reads what you have selected in the running app.

What it inherits by attaching

Your Figma seat, and with it a hard quota rather than a permission boundary. Starter View and Collab seats get six tool calls a month; Professional Dev and Full seats get 200 a day at 10 a minute, Organization 600 a day. Per-minute limits stall bursts even when the daily budget is intact.

The only server here that touches nothing of yours — and the one whose most-copied setup names a tool removed in December 2025.

What it attaches to

A hosted documentation index at mcp.context7.com/mcp. Or nothing at all: npx ctx7 setup can install it as a skill that shells out to ctx7 library and ctx7 docs, with no MCP server in the loop.

What it inherits by attaching

No access to your code, your account or your machine — it is a read path into someone else's documentation. What it costs instead is context: the CLI-plus-skill mode exists precisely so tool schemas are not loaded into every request.