playwright mcp in github copilot
Copilot's agent mode can call MCP tools, which means it can drive a browser if you give it one. The setup is short. What is worth reading is the part after the setup, because an editor is a different environment from a chat client and two of the defaults that are fine elsewhere are wrong here.
Copilot reads MCP servers from a workspace file, .vscode/mcp.json, or from
your user settings if you want it in every project. The block is the same shape
every client uses: a command and its arguments.
For Microsoft's server:
{
"servers": {
"playwright": {
"command": "npx",
"args": ["@playwright/mcp@latest", "--isolated"]
}
}
}For this project's server the block is in the MCP server, which also lists the settings and every tool with what it returns.
Then switch Copilot Chat to agent mode. Ask mode does not call tools, and this is the single most common reason people conclude the server is not working when it is registered correctly.
The example above passes --isolated on purpose. In an editor you are very
likely to be the person who also has an MCP browser registered in a terminal
assistant, and two clients on one persistent profile produce
browser is already in use. An editor
is exactly where that collision shows up, because an editor is where people run
two assistants at once without noticing.
If you need to be logged in, prefer --storage-state over a shared persistent
profile, for the same reason.
A browser server adds twenty-odd tool descriptions to the model's context on every turn, and in an editor you probably already have several servers registered. Copilot has a ceiling on how many tools it will offer at once, and crossing it means some of them stop being available without a message that says so.
The size of one browser server, measured on ours on 2026-09-10: 24 tools, 14,064 characters, about 3,500 tokens per turn. Register two browser servers in an editor that already has a file server and a search server and you are spending a meaningful slice of the context window on tool descriptions before your question arrives.
Two habits keep this from biting:
Enable capabilities you use, not all of them. --caps is additive on
Microsoft's server. Vision, pdf and devtools off unless a page forces it.
Do not register two browser servers. Overlapping tool names double the context and make the model's choice worse. Pick one.
In an editor the model has your repository in context, which is powerful and also the source of the main failure mode: it will happily read the code and guess what the page must look like rather than opening it.
- Say "open the page" explicitly when you want it observed rather than inferred. "Check whether the login form submits" reads code. "Open localhost:3000, fill the login form and tell me what happens" uses the browser.
- Give it a stopping condition. Same as anywhere else, and worse here, because an editor session is long and nobody is watching token spend. Playwright MCP best practices has the rest.
- Point it at your dev server, not production. Obvious, routinely ignored, and the reason an agent once filled in a real form.
An editor already has the code. If the question is answerable from the source, the browser is a slow way to answer it, and every browser step costs a round trip. Reach for it when the question is about what the running page does - a layout that breaks, a request that fires, a flow that only fails in a real browser - and not when it is about what the code says.
And if what you want is a repeatable check rather than an answer today, write a Playwright test instead and let the CLI run it. Playwright MCP vs the CLI has the split.
Does Copilot support MCP? Yes, in agent mode. Ask mode does not call tools.
Where does the config go? .vscode/mcp.json for one workspace, user
settings for all of them.
Why does Copilot not use the browser tools? Almost always: not in agent mode, or the tool ceiling is hit because several servers are registered.
Can I use this with a local dev server? Yes, and it is the best use of it in an editor.
Does it work in Cursor, Cline, Windsurf? Yes, with a different top-level key in the config file. The MCP server lists which key each client expects, including the one that is not JSON.
See also: Playwright MCP best practices, running AIHawk with Claude Code, and using an agent to test your own site.
-
microsoft/playwright-mcp, retrieved 2026-09-10, for the package name,
--isolated,--storage-stateand--caps. - GitHub's own documentation for Copilot agent mode and MCP configuration.
Written while maintaining a competing MCP browser server. The example block registers Microsoft's, because in an editor it is the one most readers already have and the advice does not depend on which you chose.