MENU
Ask Claude to compare three products and it will tell you what it remembers. Ask it what those products say on their websites today and it hits a wall: it needs a search API, a scraper, and an account with each of them.
That wall is why most AI assistants stop at advice. The work they can't do isn't hard — it's just behind a paywall, and nobody has given the assistant a wallet.
Today that changes. Connect the Nevermined Catalog to Claude, ChatGPT or LangSmith Fleet, approve one budget on one screen, and your assistant can discover and pay for real services, per call, as it works. No API keys. No accounts with each vendor. Every payment on the record.
In Claude you add a custom connector and paste one URL. You leave the OAuth client ID and secret empty — there is nothing to register. In ChatGPT it's a custom MCP app. In LangSmith Fleet you enter the endpoints once.
Then you just ask for something.
The first time your assistant needs to spend, it doesn't ask for a key. It sends you to Nevermined to approve a budget: who's asking, how much they can spend, and for how long. We set $5 for 7 days. You click Grant access, and the assistant picks up exactly where it left off.
That approval is the credential. The assistant never holds a key — it holds five dollars, for a week, spendable only through the Router, revocable in one click.
We recorded three runs on production. Real money, real services, no test merchants.
Claude was asked which browsers are making the loudest privacy claims right now, and to fetch what each says on its live product page. It found them, paid for a search, paid to fetch three pages, and wrote the comparison itself. Four paid calls. 1.6 cents.
ChatGPT planned a long weekend in Lisbon: the actual forecast from a weather API, family-friendly restaurants from a places API. Two paid calls. 1 cent.
A LangSmith Fleet agent produced a one-page company brief from a live website — the thing a salesperson does before a call. Two paid calls. 1.2 cents.
Three assistants. Three budgets. Under four cents in total, and five different vendors paid — none of which we told the assistants to use, and none of which we hold an account with.
Something happened in testing that's worth being honest about.
Our first prompt asked Claude for a competitive teardown — and it charged us nothing. It answered from what it already knew, which was the right call. An agent that spends your money on something it can answer for free is a badly behaved agent.
Once we asked for what those pages say today, it paid immediately — because now it genuinely couldn't know.
That's the shape of this. Your assistant pays for the things it cannot get on its own — live data, specialist APIs, someone else's work — and does the reasoning itself, for free. You're not renting intelligence. You're buying the inputs it can't reach.
Open your account and you see the three grants: which assistant, what it may spend, what it has spent. Below that, every individual payment — the merchant, the amount, the transaction hash.
This is the part that makes the rest comfortable. An agent with a budget is only a good idea if you can see what it did with it, and stop it instantly. Disconnect revokes every credential the agent was issued; the next call it makes fails.
An API key is permanent, unscoped and invisible. Once an agent has one, it can spend until you notice.
A Nevermined grant is the opposite: capped in money, bounded in time, scoped to one thing, visible line by line, and revocable in a click. The cap is the total debited — our routing fee comes out of the same budget, not on top of it.
And it's one grant per assistant, not one account per vendor. The catalog carries hundreds of services; your assistant reaches all of them through a budget you approved once.
If you use Claude, ChatGPT or LangSmith Fleet, this takes about two minutes. Start in sandbox with testnet funds if you'd rather rehearse.
See Nevermined
in Action
Real-time payments, flexible pricing, and outcome-based monetization—all in one platform.
