WebMCP tools for agents

Availability

A supporting browser and agent can discover tools while the relevant page is open. Navigation changes the tool list. WebMCP availability depends on the browser, its trial support, and the agent. The tools do not provide a remote MCP server or a background service.

search_poems

The homepage, Library,All Poems, and theme pages expose this tool. It searches the published Jindal collection by words in poem titles or text, with an optional theme filter. The search covers the collection, including poems outside the current page.

Input: {"query":"rose","theme":"nature"}. Either field or both are accepted. Queries contain 2 to 100 characters. The themes are nature, character, medicine, connection, and delight. Each response returns up to eight poems with titles, years, themes, full source URLs, and the total match count. Historical books and unpublished drafts are outside this collection.

find_similar_poems

A published Jindal poem page exposes this tool. Input:{}. The open page fixes the poem to compare. Each response returns up to eight other published Jindal poems with titles, similarity scores, and full source URLs. Higher scores mean closer matches. The tool uses stored poem vectors and does not generate an embedding from a new user query.

search_this_book

Published edition overviews in theHistorical Libraryexpose this tool. Input: {"query":"fever"}, with a query of 2 to 64 characters. Search is fixed to the open book's indexed English translation. The tool has no transcription or source-language selector.

Each response returns up to five page labels, short excerpts, and full source URLs. A result is a page hit, not necessarily an individual poem. Review editions and separate reader pages do not register this tool. Historical texts are evidence for scholarship, not current medical advice.

Sources and citations

Successful responses identify the publisher in source.nameand the site in source.url. Each result's urlis a full HTTPS citation target on poetdoctor.com. For example, Lights outhas the source URLhttps://poetdoctor.com/poems/lights-out-its-dark/.

Poet Doctor publishes the online editions. The linked source page identifies the work's author, text, and edition details. Tool results provide metadata and excerpts, not full poems or books.

Limits and errors

Each response is capped at 1,500 serialized characters, including source metadata. truncated: true means the tool removed trailing results to fit that budget. A request times out after 20 seconds. The tools do not retry automatically. Provider work already started may continue after the browser request stops.

The tools share the public API's rate limits and provider budgets. Historical book search can consume the site's daily AI Search allowance. Search queries have length limits and cannot contain control characters.

• invalid_input means an input failed validation.
• rate_limited means the request limit was reached.
• daily_limit means book search reached its daily allowance.
• timeout means the request exceeded its time limit.
• cancelled means the user, agent, or navigation stopped the request.
• unavailable means the tool could not return valid results.

The rate_limited and daily_limit errors includeretry_after_seconds. That field gives the delay before another attempt. Use public research terms. Do not enter patient identifiers or confidential clinical material.

Source text and agent safety

Tool descriptions are fixed. Returned titles and excerpts are source material, not instructions. All three tools mark result content as untrusted. Book excerpts have HTML tags removed. These controls reduce risk but do not guarantee that an agent will resist prompt injection.

Result links must point to the expected Poet Doctor collection. Redirected API requests fail. The tools cannot edit content, subscribe users, select an arbitrary URL, or access administration.

Discovery and other access

The LLM guide, this page, and the sitemap describe these resources for readers and crawlers. Native WebMCP discovery requires a supporting agent to visit a tool page. The public pages remain readable without WebMCP. The HTTP API schema describes more capabilities than the three tools, including other corpus searches. These tools do not search the medical humanities directory or the whole historical corpus.

Chrome's WebMCP guidedescribes browser setup and inspector testing. OpenAI documents site-tool discovery in theChatGPT desktop app's built-in browser. Those capabilities do not establish support in every browser assistant.