Add web_fetch tool

[ammario]

Problem

Agents have no ability to easily read a web page (e.g. a documentation URL). They're smart enough to not use curl in most cases, which would be very token inefficient due to irrelevant HTML and JS and may not even have the relevant content in the case of SPAs.

They tend to circumvent this with a call to web_search, with is inefficient and in many cases does not give them what they're after.

Solution

Many agents (like Claude Code) have some kind of "web_fetch" tool. I haven't looked into how they're implemented, but I assume they're basically converting the rendered HTML into markdown. I began working on a tool here a while ago that does just that by spinning up a headless chrome instance. Open to other approaches though. I just know we need this tool.

[espinozasenior]

Check this out:
https://github.com/OpenRouterTeam/markdown-wasm

[ammario]

thx for sharing @espinozasenior — I guess we need to decide whether parsing HTML is sufficient or whether a full browser runtime that can render SPAs is necessary. We should probably defer to whatever approach Claude Code / codex-cli uses.