agentweb

AgentWeb lets your coding agent read and act on real websites — GitHub,
Amazon, Spotify, Wikipedia, and more — with a plain command instead of
clicking around a browser page.

Reads are the easy part; a search engine can do that. The reason to use
AgentWeb is the other half: it can add to a cart, post a comment, or follow
an artist on an account that's actually yours, and it tells you exactly
what it's about to do before it does it.

Install

pipx install agentweb-cli   # or: pip install agentweb-cli

Want it wired into Claude Code or Codex automatically too? Use the install
script instead of pip:

curl -fsSL https://github.com/AnayGarodia/agentweb/releases/latest/download/install.sh | sh

Windows: irm https://github.com/AnayGarodia/agentweb/releases/latest/download/install.ps1 | iex,
or pip install agentweb-cli works there too. Details, opt-outs, and every
other agent host: agent & MCP setup.

Watch it act

agentweb connect news.ycombinator.com                          # once: sign in, in a real browser window
agentweb hn favorite --item-id 8863 --dry-run                  # preview only, nothing changes yet
agentweb hn favorite --item-id 8863 --confirm                  # does it, then re-reads the site to prove it
agentweb hn favorite --item-id 8863 --confirm --enabled false  # undone, the same way

The login is a one-time step, and it's yours: a real browser window opens,
you sign in, and AgentWeb keeps only that site's session afterward — never
your password. Every command after that is typed, not clicked.

Try a read yourself, no login needed

agentweb find --run "find the latest version of React on npm"
agentweb npm get-version --package react

The first line is a plain-language goal; find --run resolves it to the one
unambiguous read it names and executes it immediately, no login or guesswork
needed. The second line is the literal typed command underneath it — same
site, same answer, same real live JSON, reached the other way. Use whichever
fits: a goal when you don't know the command, the command directly once you
do.

AgentWeb vs. a browser agent

A browser-driving agent gets a live, cookie-bearing browser and has to
re-read and re-guess the page on every single run. AgentWeb maps a site's
commands once, so your agent calls hn favorite directly instead of
finding and clicking a button, and it can't accidentally do anything else
on the page. Browser automation is still the right call for a site nobody
has mapped yet.

What it can do

Where things stand

Coverage varies by site: some are fully mapped, others cover the common
tasks and say plainly what's left. Independent double-checking of every
write is being built next — today AgentWeb trusts a successful response
plus, where the adapter supports it, a live re-read like the one in the gif
above. A few sites also block automated traffic and need the browser-login
fallback in security and trust. Nothing here is hidden:
agentweb sites and agentweb capabilities DOMAIN show the real state of
what's installed on your machine, right now.

Open source, closed certification

This repository — runtime, CLI, MCP server, and the full adapter catalog
above — is MIT-licensed: read it, fork it, submit an adapter for a new site.
What's closed is the toolchain that turns a working adapter into a
certified one (crawler, contract/evidence checks, release gates, registry
signing) — a contributed adapter is honestly labelled community
(uncertified) until that review signs it, and every interface (agentweb sites, agentweb describe, and both MCP tools) says which one you have. See
trust labels for the full mechanism, or
build an adapter to contribute one.

Learn more

Docs · Supported sites · Agent & MCP setup · Contributing · Security