Web

Fetches a URL and hands the agent the article — navigation, banners and comment threads stripped — with the page title and the address it ended up at after redirects. A page too long for one reply is paged rather than cut off, and read again from memory rather than downloaded twice — and it can be indexed first, by heading or by pattern, so the agent reads the part that answers the question instead of the first part. Search covers Wikipedia, Stack Overflow, crates.io and npm — the indexes that need no account and beat a general engine at their own subject — and the open web for everything else, with every result saying which one answered. General search is rationed rather than unlimited, because every keyless engine is; a burst of searches falls back to a smaller independent index and says so. http and https only, and only the public internet: anything resolving to your own machine or network is refused, on the first request and on every redirect, so a page cannot send Gantry looking at your router. No key, no account, nothing to pay for.

What it can do

Examples of the tools the agent gets from this connector. Each call shows up in the activity feed with its input and result.

1. websearch “cannot borrow as mutable”
2. webfetch_url docs.rs/tokio
3. webfind_in_page a 300,000-character spec

How it connects

Sign-in

No sign-in needed. No account and no key. Install it and it is ready.

Runs as

Built into Gantry. Part of Gantry itself, written in Rust. No runtime to install.

Permissions

Reads run in every mode. Writes outside your folders, messages and anything destructive follow the chat's permission mode: they ask in Manual and Auto-edit, and are judged by the guard model in Auto. How permissions work

Install in Gantry

1. 1

   Open Connectors in the sidebar. The catalog ships with the app; nothing is downloaded until you ask.

2. 2

   Find Web and choose Install. That is all.

3. 3

   Attach it to a chat, or leave it available to every chat in a project. The agent sees its tools from then on.