Give your agent the web.
Keep the receipts.
AgentBrowse reads allowed public webpages and returns clean text, the source URL, and short quotes you can verify. Start with five free reads every day. Use it through REST or MCP with your existing Agent Media Tools key.
Your task cannot wander beyond the sites you explicitly list.
Use one REST request or the browse_web MCP tool.
Receive the title, URL, outcome, source quotes, and a stable SHA-256 for each text excerpt.
Verified on live sites — not just a demo
Public catalog suite (server HTTP reader): 28/28 hard requirements on the latest run, including 2/2 private-network blocks that must never succeed. Soft outcomes (JS walls, timeouts) return clear codes — we don’t invent page content.
-
22
Useful extracts — real titles/paragraphs from public HTML pages in the catalog
-
safety
SSRF blocked — localhost and cloud metadata IPs rejected before fetch
-
honest
Methods public — re-run
npm run test:sitesin the runtime repo
Full snapshot, limits, and “what we don’t claim”: AgentBrowse verification →
How it works
- Your agent asks about a public page — for example “What does this pricing page say?”
- Who opens the page?
- With Companion — the page opens from your computer and home/office internet (more private).
- Without Companion — we open the public page on our side when we have room.
- You get proof back — short quotes, title, and URL so you can verify what was seen.
- Safety first — each task lists which sites it may touch. Private/home networks and cloud admin addresses are blocked so agents can’t wander where they shouldn’t.
A hybrid reader—not an expensive browser farm
AgentBrowse picks the smallest safe execution path for each read. The contract stays the same, so your agent does not need separate integrations for hosted and local work.
| Path | Best for | Trust boundary |
|---|---|---|
http_read | Docs, articles, product pages | Runs on our bounded reader |
local_browser | Public pages better reached from your network | Runs through your verified Companion |
cloud_browser | Operator-enabled JavaScript rendering | Isolated server Chromium; higher cost |
One safety policy everywhere: hosted and Companion reads both use DNS-pinned connections, re-check every redirect, enforce the task’s domain allowlist, cap response size and time, and block private or metadata networks. Companion authenticates over encrypted WebSocket with your normal AMT key; the runtime maps it to your account without storing the key. Every text-evidence item includes content_sha256 so identical source excerpts can be compared reliably across runs (an integrity identifier, not a publisher signature).
Download Companion 0.4.0 beta
Local by design: allowed public pages are downloaded from your home or office internet instead of a shared cloud worker. Companion returns structured text; it is not a remote desktop or a signed-in browser session. One file per OS, no Node install.
-
run
Linux / macOS
chmod +x agentbrowse-companion-*then run it. Windows: double-click the.exe(SmartScreen may warn on unsigned builds). -
connect
Paste key → Connect Control panel opens at
http://127.0.0.1:19556. The hosted bridge is already filled in; use your normal AMT API key. -
status
Confirm it’s working Green “Ready” in the Companion panel, or call
GET /api/browser/companionwith your Bearer key — should showcompanion_online: true. Also shown on the Agent Hub.
Binaries: release details · Verify the published SHA-256 checksum before running · From source: npm run companion:tray · Beta builds are not code-signed or notarized yet (Gatekeeper/SmartScreen may prompt). Older beta binaries used a local-only bridge address; download the current build for hosted pairing.
When a page needs more than HTTP
Our default server path is a light public-page reader. Client-heavy apps behave differently:
- successSuccessful HTTP read — title, quotes, and useful text extracted. Good for docs, blogs, many news and product pages.
- partialPartial content — something came back but it’s thin (
content_quality: partial/partially_completed). Worth checking the quote. - js wallHard failure — JS wall, bot challenge, or timeout (
read_outcome: js_wall | blocked | timeout). Server HTTP is not enough; use Companion so the page loads on your network (or cloud browser if your operator enabled it).
Mode auto tries Companion first when online, then server HTTP, then asks for Companion again on JS walls — you rarely need to pick a mode manually.
Free to start. Clear after that.
No surprise product plan. Everyday page reads include a free daily allowance. After that, successful reads use one flexible credit — the same pool as screenshots and other light tools.
-
free
5 light reads every day No card required. Paid plans receive a higher daily allowance.
-
then
One credit per successful read Whether we open the page on our side or Companion opens it on yours. Same balance as the rest of Agent Media Tools.
-
shared
One wallet No separate AgentBrowse subscription. Buy credit packs or use Founding Pro monthly credits like everywhere else.
Why charge after free? Free covers exploration. After that, a small credit pays for login checks, keeping track of each job, saving results, and reading pages on our servers when we have room. If our servers are busy — or you use Companion — the heavy lifting moves to your computer. Shared capacity is limited; light metering keeps free use fair for everyone. Full details: Pricing.
For developers
Recommended default: mode: "auto" — Companion if online, else capacity-gated server HTTP; on JS walls / bot blocks, falls back to Companion (or waits until you connect). REST POST /api/browser/browse and MCP browse_web share the same body fields and response shape.
| Mode | What it does | Usage |
|---|---|---|
auto ★ | Default. Companion → HTTP → Companion on JS/bot walls | As used |
http_read | Server public HTML only (no Companion rewrite unless local-first + device online) | Free daily → 1 credit |
local_browser | Always Companion; parks as waiting_for_local_browser if offline | Companion required |
cloud_browser | Server Chromium (operator opt-in only) | 3 credits when enabled |
evidence: "short" (default) or "paragraph" for longer research quotes. Companion status: GET /api/browser/companion.
| Method | Path | |
|---|---|---|
| POST | /api/browser/browse | Sync browse |
| POST | /api/browser/tasks | Async task |
| GET | /api/browser/tasks/:id | Poll result |
| MCP | browse_web | Same payload |
Safety: SSRF blocks for private IPs and cloud metadata; allowed_domains required. Capacity: server HTTP sheds under RAM/load/slot pressure (GET /v1/policy on the runtime). Local bridge: WS /bridge/ws.
Example request
After free daily reads, a successful page read uses one flexible credit.
curl -X POST https://agentmediatools.com/api/browser/browse \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"goal": "What is this page about?",
"start_url": "https://example.com",
"allowed_domains": ["example.com"],
"mode": "auto"
}'
Common questions
Do I need a separate product key?
No. Use your normal Agent Media Tools API key. Usage shares the same flexible credit balance as other tools.
What does this actually do?
It lets an AI agent read public webpages and return answers with short quotes and links you can verify. The optional Companion downloads allowed pages from your computer instead of ours.
Is my browsing private?
With Companion connected: allowed public-page downloads use your device and network. We still coordinate the request and receive the structured result. Without Companion: public pages are read through our service. Private and home-network addresses remain blocked either way.
Why not use the web access my agent already has?
Built-in model browse is fine for casual chat, but weak on proof and control. DIY fetch/Playwright means you run the infrastructure and safety risk. AgentBrowse gives allowlisted public reads, blocked private networks, quotes you can verify, five free daily reads, and an optional local Companion under the same key as your other tools.
How is this different from structured-scrape?
Structured scrape fills a fixed list of fields (like “price”, “title”). AgentBrowse is better when you want research-style answers with proof quotes, not a rigid form.
What if a site blocks bots?
You’ll get a clear message. Turning on Companion often helps (real user network). If we charged you and our service was down, that usage is refunded.
Can it open my Gmail?
No. The current Companion is a public-page reader, not control of your signed-in Chrome session. Logged-in browser support would require a future extension and is not available today.
What’s all the developer jargon?
Modes like auto / http_read / local_browser, SSRF protection, MCP browse_web, and capacity gates are for builders — see For developers. Everyday use only needs an API key and (optionally) Companion.
Give your agent the public web
Start on the free daily allowance. Add Companion when you want pages to open from your computer.
Related: Structured scrape for fixed field extraction · Pricing · Agent Hub for keys and MCP.