AgentBrowse · verification

Does it actually work on the real web?

We don’t claim to beat chat “browse” buttons. We publish a live public-site suite you can re-run: real extracts when pages allow it, clear codes when they don’t, and hard fails for private-network attacks.

Try free daily reads → Product Re-run

Last run: 2026-07-25
Runtime: 0.3.1
Mode: server http_read
28 targets

28/28 Suite passed
(hard checks)

22 Useful page content
extracted

4 Clear soft outcomes
(JS / bot / timeout)

2/2 SSRF attacks blocked
(must never succeed)

How to read this: 28/28 means every required success and every required security block passed. 22 of 28 returned usable page text; the other 4 public targets returned structured failure codes (not invented content). 2 targets are deliberate SSRF attacks that must fail.

Sample of this run

Target	Outcome	~ms	Notes
example.com	success	223	Title + content
Wikipedia / MDN / GitHub Docs	success	160–340	Useful public HTML
BBC News	success	386	News home
reddit.com	javascript_required	138	Clear code → Companion path
npmjs.com	AUTHENTICATION_REQUIRED	114	403 handled, not invented
NPR	NAVIGATION_TIMEOUT	~25s	Slow/blocked from suite host
127.0.0.1 / cloud metadata	PRIVATE_NETWORK_BLOCKED	0	Must never open

Full evidence: 28 cases in the runtime repo (tests/sites/catalog.ts). Download this run’s machine-readable results → Soft sites pass the suite when the tool reports a clear reason instead of fake text.

Safety

Private networks — 127.0.0.1 and cloud metadata IPs are blocked before fetch so agents can’t pivot into your LAN or cloud admin endpoints.

Site allowlists — each task lists which domains it may touch. Safety is product policy, not hope.

What we don’t claim

Not a leaderboard vs ChatGPT/Claude browse · Not 100% of the internet (JS/login walls need Companion) · Not a permanent score (sites change; this is a dated snapshot).

Re-run

cd agentbrowse
npm install
npm run test:sites -- --json ./agentbrowse-benchmark.json

Results vary by network path (home vs datacenter). Datacenter IPs see more bot walls — the suite expects clear codes, not silent garbage.

Try free daily reads → Why AgentBrowse Pricing

Free daily reads cover real public pages (many catalog sites return useful text). After free, one light credit pays for coordination and server HTTP when capacity allows. Hard pages get a structured failure or Companion — not a hallucinated summary from empty HTML. Safety blocks are in the suite, not just the marketing copy.