Skip to main content
The Scrinly skill teaches compatible coding agents how to use the remote MCP server safely. It selects the correct tool, calculates the smallest valid maxCredits ceiling, preserves direct assets, polls durable jobs without resubmitting them, and reports actual refunds and net credits. The skill contains no credentials and does not replace the MCP connection. Configure the MCP server with the recommended OAuth flow first, then install the skill. Server-to-server clients may use the documented API-key fallback instead.

​
Install

With Node.js 20.12 or newer, install the Scrinly skill globally from its public repository:
The global installation makes the skill available across projects for agents that discover the shared .agents/skills location. To keep it in only the current project, omit -g. You can inspect the source before installing it at github.com/davmixcool/skills.
Agent skill discovery varies by client. Claude Code users may alternatively copy the skills/scrinly directory into ~/.claude/skills/. Restart an already-running client after installing or updating a skill.

​
What the skill enforces

  • Billable capture and comparison tools run only when requested or clearly authorized.
  • maxCredits matches the requested screenshot add-ons instead of using a broad ceiling.
  • A returned jobId is polled with get_job_status; the original billable call is never silently repeated.
  • A screenshot stays paired with its matching region manifest and hashes.
  • Partial region or Visual Style Guide failures preserve successful screenshot evidence and report the component refund.
  • Provider keys, page credentials, storage credentials, and webhook secrets never enter MCP tool arguments.
  • Direct provider URLs remain resource links rather than being proxied or converted to base64.
OAuth permissions still govern which tools the connected agent can call. The skill cannot expand its scopes, bypass Scrinly’s maxCredits guard, or access an account other than the one approved during authorization.

​
Example requests

Check the connection and allowance without spending credits:
Capture one stored screenshot:
Capture a full page as model-sized regions:
Generate a Visual Style Guide:
For asynchronous work, keep the job ID in the same conversation:

​
Results and failures

A successful capture returns concise text, structured JSON, and resource links for the complete screenshot and any region crops. A comparison may also return an overlay, heatmap, or mask link. Report the terminal status and the core API’s actual credits.charged, credits.refunded, and credits.net values. If an add-on fails but the screenshot succeeds, the skill treats the result as usable partial evidence. If a full-page capture is unbounded or incomplete, or a diff manifest does not match its screenshot hash, the skill reports the typed error and does not weaken validation or spend credits on an automatic retry. The skill source includes detailed tool contracts and workflow references. The public API remains the authority for validation, billing, storage, and refunds.