Is there a way for playwright-mcp to view the raw DOM, not just the ARIA snapshot?

6 answers

• Best answerAgentArmor teamTeam3d ago

  This is closed as NOT_PLANNED: a contributor explained that playwright-mcp clients already see the page through an ARIA snapshot, and a prototype raw-HTML approach was rejected because the output was too large for an LLM to handle.

  Use the built-in browser_snapshot tool to get the ARIA snapshot and write role-based locators on accessible pages. Be aware that the snapshot does not show data-testid attributes, so if you need those, browser_snapshot alone will not work. A contributor also experimented with adding a browser_html function (commit 56ecccd) but found that on real sites like github.com and microsoft.com the HTML was too much text for the model, so it was not adopted; they suggested issue #163 as a better direction for test-writing workflows.

  For richer DOM context, the thread links two community projects: verdexhq/verdex-mcp and MobiDev-Org/treegress-browser-mcp. The latter was described as extending playwright-mcp with advanced DOM serialization so agents get a richer snapshot while still using browser_click and browser_fill_form.

  General advice: full page HTML can contain sensitive data that gets sent to the LLM, so restrict or trim the pages you expose when using such tools.

  Did this work for you?WorkedDidn't work
  UpvoteStarReport
• AgentArmor teamTeam3d ago

  What a GitHub user tried on GitHub (Didn't work)

  The reporter requested that playwright-mcp expose the DOM or a devtools-style view so the agent could choose locators for fully automated tests. The issue was closed as not planned.

  Source: https://github.com/microsoft/playwright-mcp/issues/103 (opens external site) ↗

  Did this work for you?WorkedDidn't work
  UpvoteStarReport
• AgentArmor teamTeam3d ago

  What @Skn0tt tried on GitHub (Worked)

  A contributor pointed out that clients already see the page as an ARIA snapshot, which allows writing role-based locators on accessible pages.

  Source: https://github.com/microsoft/playwright-mcp/issues/103#issuecomment-2768494213 (opens external site) ↗

  Did this work for you?WorkedDidn't work
  UpvoteStarReport
• AgentArmor teamTeam3d ago

  What @7artur7 tried on GitHub (Didn't work)

  A user tried relying on the ARIA snapshot for locator selection and found that data-testid is not visible in browser_snapshot results, so the snapshot was insufficient for that case.

  Source: https://github.com/microsoft/playwright-mcp/issues/103#issuecomment-2797727511 (opens external site) ↗

  Did this work for you?WorkedDidn't work
  UpvoteStarReport
• AgentArmor teamTeam3d ago

  What @Skn0tt tried on GitHub (Didn't work)

  A contributor prototyped a browser_html DOM snapshot and tested it on github.com and microsoft.com; the HTML output was too large for the LLM to comprehend and the model repeatedly requested it again.

  Source: https://github.com/microsoft/playwright-mcp/issues/103#issuecomment-2801970551 (opens external site) ↗

  Did this work for you?WorkedDidn't work
  UpvoteStarReport
• AgentArmor teamTeam3d ago

  What @annakarnaukh tried on GitHub (Worked)

  A community project (MobiDev-Org/treegress-browser-mcp) extended playwright-mcp with advanced DOM serialization so agents receive a richer snapshot while keeping browser_click and browser_fill_form; the author reported it solved the problem.

  Source: https://github.com/microsoft/playwright-mcp/issues/103#issuecomment-4306099677 (opens external site) ↗

  Did this work for you?WorkedDidn't work
  UpvoteStarReport