Remote URL policy — anymd

anymd accepts a url source because many agents are handed a PDF link rather than a path. Remote URLs are in scope, but only as the narrow fetch described here.

What a url source is for

  • One http(s) document body, fetched by the local process and treated exactly like a local file after download.
  • An explicit convenience for a caller that already has a link.

It is not a browser, a crawler, a general web client, an authentication delegate, or a way to reach a private network.

What the fetch does

  • http and https only; every other scheme is rejected.
  • The hostname is resolved locally and the resolved address is pinned for the connection: the address the guard checked is the address the socket connects to. A fresh, pool-free client makes every hop a separate validation and connection boundary.
  • Redirects are followed only after the target is resolved and re-validated, up to five hops.
  • Non-public addresses are rejected by default.
  • MCP_PDF_ALLOW_PRIVATE_IPS=true is an explicit opt-in for local fixtures or a trusted network; it is never the default.
  • A 30-second timeout, a 256 MiB body limit, no environment proxy, and no file:// access.
  • The downloaded body is written to a secure temporary file and removed after use.

Tests

crates/anymd-core/src/url_fetch.rs covers the guarantee, including one_dns_resolution_is_pinned_to_the_actual_connection and the redirect re-validation tests.

Boundaries

A url source exists so an agent with a single link does not have to fetch it first. It is not a web research tool.