Fetch web page
Let the LLM read a web page by URL, as plain text with the page’s scripts, styling and navigation stripped, one page of text at a time. Pair it with Search the web, which finds the URLs.
Ingredient Family: llmQueryToolIngredients
Outline Display: Allow LLM to read web pages ({{toolName}}): {{toolDescription}}
Characters per page: {{charactersPerPage}}
Bytes read per page: {{maxBytes}}{{domains:
Domains: {{domains}}
Listed domains are blocked: {{domainsBlocked}}}}
Requires approval: {{requiresApproval}}{{advisory:
Not enforced by the server: {{advisory}}}}
Details
The model sends a url; the server fetches it and answers the page’s text, charactersPerPage characters at a time with a nextCursor for the rest, which the model sends back alone as cursor to keep reading. Only https pages are read, at most maxBytes of each (a page cut there says so in its answer), and HTML comes back as text – scripts, styles and the nav, header, footer and aside chrome around the main content dropped, block elements on their own lines, entities decoded. Plain text, Markdown, JSON and XML (feeds included) come back as they are; any other content type, PDFs included, is refused with a message naming the type, and a page served with no type is read as HTML when it opens with a tag, as text when its bytes are UTF-8, and refused otherwise. A page in another encoding, declared by its Content-Type header or its own <meta charset>, is read as UTF-8, with anything the declared encoding cannot decode replaced by U+FFFD; one in an encoding the server does not know, or in UTF-16 or UTF-32, is refused. Up to three redirects are followed, each judged like the URL that led to it, and the continuations of one chat turn read the text the first call fetched rather than fetching the page again.
domains is a comma-separated list of domain names, each covering the domain itself and every host under it (example.com covers docs.example.com): the only domains whose pages may be read, or, with domainsBlocked: true, the domains whose pages are never read. A URL outside the policy is refused before anything is fetched, and an entry that names no host – a scheme the app may not reach, say – fails every call, naming the entry, until the setting is corrected. Whatever the policy says, the server never connects to its own network: a host that resolves to a loopback, private, link-local or otherwise non-public address is refused at connection time, so a page cannot be used to read the server’s neighbours.
Every call is audited under the tool’s wire name with the URL it read as its arguments, continuation calls included. A url that is not an absolute https URL, or a page the site does not serve (404, 410), fails the call as bad arguments, a refused or failing site as a malfunction, and a URL outside the policy as a refusal; the model is told which.
Settings
Required
toolName: constant value (type: text; required, non-nullable)
The name the LLM calls this tool by: its wire name, which MCP clients, approvals, audit lines and dashboard policy all key on, so renaming it is renaming the tool for every one of them. Required. 1-64 characters of letters, digits, underscore and hyphen, taken exactly as written; a tool whose companions append a suffix to it (a lookup’s_distinct_values, a SQL query’s_schema) must leave room for the suffix under the 64. Two tools in one module cannot share a name.
Optional
-
toolTitle: constant value (type: text; optional)
The human-readable name an MCP client shows for this tool (MCP’stitle), never read by a model. Omitted, it is derived from the wire name:search_open_ordersbecomes “Search Open Orders”. Two tools in one module cannot share a title, derived or authored. -
toolDescription: constant value (type: text; required, non-nullable)
What the LLM should read web pages for, to help it decide when to open a page rather than rely on a search snippet -
advisory: constant value (type: text; optional)
A rule the server does NOT enforce, sent to the model as guidance under an explicit ‘not enforced’ header; use only when no deterministic form (Require, parameter constraints, allowedRecipients…) exists. The checker warns and the dashboard shows it as not guaranteed. -
requiresApproval: constant value (type: bool; required, non-nullable)
Require the user to approve each invocation of this tool before it runs. Only honored inside a Chat concept, where the agent pauses mid-turn and shows Approve/Reject buttons; rejecting tells the LLM the call was denied and ends the turn. Setting it true on a tool used by any other agent (e.g. a row-action AskAnLLM triggered by a button) is rejected by the NectryCore typechecker – except under MCP Server, which only warns. When unset, the tool is gated exactly when its steps make it anything but read-only – the same reading its MCP category comes from, so a tool that annotates itself read-only is not gated – and only where a pause is honored: an unattended surface never invents a gate it cannot keep. -
domains: constant value (type: text; optional)
Comma-separated domains, each covering the hosts under it: the only domains whose pages may be read. An entry may be spelled as a URL, or with a port or path, of which only the host counts. Unset, any https page may be read. -
domainsBlocked: constant value (type: bool; required, non-nullable)
When true,domainslists the domains whose pages are never read instead, and any other https page may be read. -
maxBytes: constant value (type: integer; required, non-nullable)
The most bytes of one page that are read; a longer page is cut there and the model told so. From 1000 to 5000000. -
charactersPerPage: constant value (type: integer; required, non-nullable)
Characters of text one call returns; a longer text comes back a page at a time through a cursor. From 500 to 50000.