Split screenshot into its own read-only tool

[masnwilliams]

Summary

screenshot was a terminal action inside computer_action. That bundled the one browser capability every multimodal model can use with the one that only works on models with pixel grounding, and it put a read-only capture behind a tool marked destructiveHint: true — so clients with per-tool approval prompt on taking a screenshot.

This splits screenshot out as a standalone tool and repositions the remaining tools:

• screenshot — new, read-only (readOnlyHint: true, idempotentHint: true), optional region crop. Registered as its own toolset so it can be enabled/disabled independently of computer.
• execute_playwright_code — description now states it's the primary way to drive a browser, and points at ariaSnapshot() for locating elements without an image.
• computer_action — no longer handles screenshots. Description reframes it as the fallback for surfaces a selector can't reach (canvas, embedded PDFs, native dialogs, drag), and notes that coordinate accuracy depends on the calling model's grounding.
• next_actions hints from manage_browsers and manage_browser_pools now suggest execute_playwright_code + screenshot instead of computer_action.

Bug fixed along the way

Region screenshots returned a cropped image but reported Viewport: WxH. Use these dimensions as the coordinate space. The image origin is the crop origin, not the screen origin, so a caller converting image coordinates to screen coordinates for a subsequent click had no way to know it needed to add region.x/y. The new tool reports the crop offset and the conversion explicitly.

Also drops a redundant browsers.retrieve call — viewport dimensions are only needed for the uncropped case.

Breaking change

computer_action no longer accepts screenshot as an action type. Callers batching click, ..., screenshot in one request now need two calls. That trade is deliberate: the batching win applies to the flow we're steering away from, and keeping a duplicate screenshot path in both tools would leave the destructive-hint problem half-fixed. Clients re-read the tool list on connect, so this lands on reconnect.

Not included

No screenshot downscaling. Pixel-grounding accuracy degrades above roughly 1024px wide, and the usual fix is to downscale server-side and rescale returned coordinates back up. That changes coordinate semantics for computer_action and is worth doing separately if we keep investing in the coordinate path.

Testing

tsc --noEmit clean, prettier --check clean on changed files, next build compiles (full build needs Clerk/CLI env vars this checkout doesn't have). Verified tool registration by instantiating the server and listing tools: screenshot registers, computer_action no longer advertises a screenshot action, and KERNEL_MCP_DISABLED_TOOLSETS=screenshot,computer removes both. The repo has no test suite, so there are no automated tests to add or run.

---

Note

Medium Risk
Breaking MCP contract: computer_action no longer supports screenshot, so existing clients must migrate; behavior change is limited to tool surface and agent guidance, not auth or data paths.

Overview
Screenshot moves out of computer_action into a dedicated screenshot MCP tool (readOnlyHint: true, optional region crop, its own screenshot toolset). computer_action no longer accepts a screenshot action type—callers that batched clicks plus screenshot need separate tool calls.

Tool guidance is updated so execute_playwright_code is described as the primary browser driver (including ariaSnapshot() for element discovery), computer_action is framed as a coordinate fallback for non-selector surfaces, and manage_browsers / manage_browser_pools next_actions point to Playwright + screenshot instead of computer_action. README and the debug prompt reference the new tool.

Region screenshots now explain crop offset vs screen coordinates (fixing misleading full-viewport wording on cropped images); full-screen captures still report viewport size. Screenshot handling is removed from computer-action.ts and implemented in src/lib/mcp/tools/screenshot.ts, registered in register.ts.

^{Reviewed by Cursor Bugbot for commit 3d56458. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
mcp	Ready	Preview	Aug 3, 2026 4:04pm

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 5fef193. Configure here.}