Wield vs Playwright MCP: what each does and when to pick it
Playwright MCP is Microsoft's MCP server. It lets an agent drive a browser through Playwright's accessibility snapshots, and the model chooses each action. Wield is also a local MCP server, but it exposes fixed, site-specific tools. Those tools run recipes with no model in the loop. Playwright MCP suits testing and sites without a recipe. Wield suits repeated tasks on supported sites.
What Playwright MCP is
Playwright MCP is an official Microsoft project under the Apache-2.0 license. You start it with npx @playwright/mcp@latest in a standard mcpServers config. It reads pages through the accessibility tree rather than screenshots, using the browser_snapshot tool. Its screenshot and coordinate tools need a model with vision.
Side-by-side
| Wield | Playwright MCP | |
|---|---|---|
| What the agent gets | Site tools such as search_connections or get_slots, built from recipes | Generic browser actions: navigate, snapshot, click, type |
| Who decides each step | A fixed recipe, with no model | The model, at every step |
| Model cost per run | $0 | Model tokens for each step |
| Browser | Your everyday browser, through the extension | A browser that Playwright controls |
| License | MIT planned for the executor, bridge, MCP server and recipe format; source not public yet | Apache-2.0 |
| Measured in our benchmark | Yes: 85% success, 0.9 s median | No |
Logins and your data
Playwright MCP uses a persistent browser profile by default, one per workspace. You can change the location with --user-data-dir. The --isolated flag starts fresh sessions, optionally with --storage-state. The --extension flag connects to an existing Chrome or Edge through the Playwright Extension. A persistent profile can serve only one browser instance at a time.
Wield uses the session you already have in your browser. It never reads cookies or passwords, and it does not ask you to paste them into the agent. Results go from the extension to wield-mcp and then to your AI app, never through wield.run.
About the benchmark
We did not measure Playwright MCP, so this page makes no speed or reliability claim about it. Our benchmark's browser-agent side ran Browser Use with Claude Sonnet 5.5, which is a different tool. As of 7 November 2026 the results are preliminary: Wield passed 183 of 208 runs and Browser Use passed 63 of 147. See the benchmark for the method and per-task rows.
When Playwright MCP is the better choice
- You test your own web app.
- You need to read or work with page layouts directly.
- The site has no Wield recipe.
- You want the model to decide each step on a page it has not seen before.
When Wield is the better choice
- You repeat the same task on a supported site.
- You want structured results that your AI app can use without parsing a page.
- You want no model tokens spent per run.
- You want to use the session already open in your browser.
FAQ
Can I use both?
Yes. Both are MCP servers, so one client can load both. Use Wield for recipe tasks and Playwright MCP for everything else.
Is Playwright MCP slower than Wield?
We have no measurement for Playwright MCP, so we cannot say. Our benchmark compared Wield with Browser Use. There, the recipe's median run took 0.9 seconds and the browser agent's took 78 seconds.
How does Wield handle writes?
Any write, such as a booking or a message, stops before its first write request. A window shows every request and its body. Deny is the default, and two minutes without an answer counts as deny.
Sources
- microsoft/playwright-mcp on GitHubgithub.com