Tools and prompts

Tools

The server exposes twenty tools to a connection (or organisation API key) with Full access; a connection or key only sees the tools its access allows. Every tool call takes a required intent: one or two sentences, 10 to 500 characters, on what the agent is trying to achieve for the user with that call and why (for example, "Screenshot the pricing page so the user can compare plans"). Urlbox records it with the call for analytics and audit, kept as long as your render history, so the agent should not put secrets in it. A call without a valid intent is refused with InvalidOptions; longer text is refused, not cut short. Connected agents are told this automatically.

Every render tool (screenshot, pdf, record_video, markdown, extract_text, scrape_html, render) takes url or html, an optional project, wait_seconds, and an options object for any other render option; call describe_options to see the full set.

• screenshot: render a PNG, JPEG or WebP screenshot of a page.
• pdf: render a page to PDF.
• record_video: record a page as a video.
• markdown: convert a page to Markdown.
• extract_text: return a page's text as Markdown plus its metadata (title, description, OpenGraph). Pass llm_prompt to also extract structured data with your project's LLM.
• scrape_html: fetch the rendered HTML of a page.
• render: the general-purpose render tool, for output formats or combinations the format-specific tools do not cover.
• get_render: look up a render by id, including one still in progress. With a project secret key, only that project's renders.
• list_renders: list recent renders, newest first, across the whole organisation or one named project. Not available with a project secret key.
• list_projects: list the projects in the organisation, marking the default project. Not available with a project secret key.
• get_project: one project's settings and default render options. Default options that hold credentials (storage keys, proxy URLs, auth headers, cookies) are listed by name only. Includes the project's public API key when the connection can render. Never returns the project's secret key: it links to the project's page in the dashboard, where you can copy it into your environment yourself.
• create_project: create a project, for example one per environment. Needs project create access.
• update_project: rename a project, enable or disable it, require signed render links, or change its engine version.
• set_project_defaults: set or clear a project's default render options.
• get_organisation: the organisation's details.
• list_webhooks and create_webhook: list a project's webhook endpoints, or add one. Signing secrets are never returned.
• report_render: report a render that does not look right, so the Urlbox team can look into it.
• usage: check the organisation's render usage against your plan for the current period. Not available with a project secret key.
• describe_options: search the render option reference by name or keyword (for example pdf margin) and return each match's description, type and allowed values, so an agent can pick the right option without guessing.

Prompts

The server also offers five prompts: ready-made instructions a user picks to start a common task. Clients show them in their own way, for example as /mcp__urlbox__screenshot in Claude Code or in Claude Desktop's prompt picker. A prompt runs nothing by itself; it hands the agent a message telling it which tools to call. Connections and keys that can render (Full access or Render and read, or a project secret key) see all five; others see none. Every argument is text, and URLs must start with http:// or https://.

• screenshot: take a screenshot, show it, and point out anything notable. Arguments: url (required); device, desktop (default) or mobile (390x844 with a mobile user agent); full_page, yes or no (default).
• compare_pages: screenshot two pages full page at the same viewport and list their layout, content and visual differences. Arguments: url_a and url_b (both required).
• save_as_pdf: render a page to PDF, return the link and summarise it in two sentences. Arguments: url (required).
• extract_article: pull the main article out of a page with a title and a three-point summary. Arguments: url (required); format, markdown (default) or text.
• mobile_check: screenshot a page at desktop (1280x800) and mobile (390x844) sizes and report responsive problems (overflow, tiny text, overlapping elements, hidden navigation), each with a severity. Arguments: url (required).

Long renders

Rendering can take longer than a single request wants to wait. Each render tool takes wait_seconds, from 0 to 95, default 60. Set it to 0 to queue the render and get its id back immediately instead of waiting at all.

If a render is still going when wait_seconds runs out, the tool returns a running status with a render id instead of the finished result. Call get_render with that id (it also takes its own wait_seconds) to pick it up once it finishes. The render keeps going in the background either way: waiting longer never loses it and never charges it twice.