unbrowseWhat if you could mine the internet?
Bitcoin pays people to burn electricity on puzzles nobody needs solved. Mining the internet would pay for something agents actually buy.
The ore is the private API behind every button.
Before the argument, the line between what runs and what doesn't, as of 12 October 2026. Learning a site's request once and replaying it is live. Paying anyone for that work is a plan: there are no payouts, no earnings and no wallets, contributor nodes run only in a staging pilot, and Unbrowse has no token.
The ore
Every website already has an API. The page you see is just its first customer.
When you search the BBC, the page doesn't search anything. It sends one request to the BBC's own search service and paints the answer.
Your agent can't see that request. So it does what you'd do:
BBC search, agent in a browser production run, 2026-09-23
open browser, load page
find the search box, type, submit
wait for results to render
read the page
──────────────────────────────────────────────
total 20.6 s
same search, one call to the BBC's route 0.8 s
Your agent spent 20 seconds doing what took the BBC under one.
It isn't one site. Same run, same kind of search:
Britannica 38.0 s → 1.7 s
BBC 20.6 s → 0.8 s
Wikipedia 17.8 s → 1.2 s
npm 11.2 s → 0.45 s
Those are the good rows. In the same run on 100 sites, a generic scripted agent got a working replay on 42. Some sites have no reusable request to find, and some block the replay.
Agents dig it up, then throw it away
Here's the strange part. Every time an agent browses a site, the request it needed flies right past it. The browser closes. The request is forgotten.
The next agent on that site digs it up again. So does the one after.
Keep it once, and every agent after can skip the browser.
We tested exactly that. We gave Claude Code nothing but Unbrowse and a fresh account, and asked it to book a flight on a demo airline site we built, one it had never seen. It used one browser session, and its browsing became a route. The next three bookings opened zero browsers.
Then it did the same on Wikipedia, which we don't control: learned the search once, then answered with no browser at all. The whole run cost $0.47.
Browse once. Every agent after calls the ore directly.
That part is live. It's how Unbrowse works today, at https://unbrowse.ai/mcp.
Nobody owns the map yet
Google indexed the web for humans and became one of the most valuable companies on earth.
Nobody has indexed it for agents.
On October 1, 2026 we probed the Tranco top 1,000: of 726 reachable sites, 37 run their own MCP server. On 362 of 425 sites we surveyed, no official API does the job agents are sent to do.
You can't draw that map from a data centre. We tried. Twenty cloud browsers cost us about $600 a month, and the bot walls still win: in a September test, even residential proxies got through on only 29 of 40 hard sites.
The fleet we want to build
The map can't be drawn from one place. It could be drawn from everywhere agents already are.
An agent working in a person's own browser is already digging. It already paid for the browser. It already passed the bot check. It's on a site someone actually needed.
So the ore would cost nothing extra to keep, and the map would grow exactly where agents ask for it.
That's mining the internet: a fleet of agent browsers that keeps what it digs up, and gets paid when it's used.
As of 12 October 2026, that fleet is our cloud browser. The pieces that would move learning onto other people's machines are built but not live in production: a browser-extension relay that lets an agent work through a person's own browser, and a contributor node that runs index jobs, which so far runs only as a staging pilot.
How the paying would work
This section is the design. As of 12 October 2026, none of it pays anyone.
Every time an agent calls a route, the proposed fee would split like this:
1 call, proposed design fee $0.010
the website $0.005
whoever dug up and fixed the route $0.004
Unbrowse $0.001
That fee is a design reference, not our price. What Unbrowse actually charges is free during the beta, 500 verified calls a month, then $10 per 10,000 verified calls after it. That is a tenth of a cent a call, and none of it is split with anyone yet.
In the design, fixing a route when the site changes earns the same way as creating it. Copying someone else's earns nothing. The split by improvement is written up in Pay for the Delta; it is not built yet.
Half goes to the website on purpose. Pay the site, and blocking agents becomes the expensive choice.
The plan pays in dollars. There is no token, and nothing to dump.
Why nobody could fake it
Nobody should be believed on their own word.
Part of that is live. The public registry today publishes a route only after a scrubbed copy, with no cookies, answers fresh queries and fails a nonsense one. Routes that need a login, a private session or a write are never public.
The network design goes further: three validators who didn't record a route replay it before it could earn anything. As of 12 October 2026, that logic is written and tested in code, and not deployed.
And it would only keep what's safe to share: public, read-only routes. Your browsing history stays yours. As of 12 October 2026, we have not built anything that records it.
Where this is
The digging works: an agent browses once in our cloud browser, and later calls can skip the browser where the route replays. That's live at https://unbrowse.ai/mcp, and setup is in the docs.
The paying isn't. As of 12 October 2026, parts of the reward rules exist as tested code that is not switched on, and we have no revenue yet. Paying customers come first, then paying the fleet, and legal review comes before any money moves.
Your agent is already digging. It's just been throwing the ore away.