Guides › For people building agents
Let your AI agent pay for APIs, without paying for broken ones
Published 21 November 2026. The setup steps below were checked against each client’s own documentation on that date; clients change their menus often, so if a step looks different, the connect page is kept current.
AI agents can now buy data on their own: one request, a few cents in USDC, no account and no API key. The hard part is not paying. It is knowing which paid API to call, and whether it will deliver once paid.
This guide connects your assistant or agent to nohumans.directory, a free directory of paid x402 APIs, so it can search by need and check an API before it pays. Then it sets up the paying side with limits, so a mistake costs cents, not your balance. It takes about five minutes, and you don’t need to write code unless you are building your own agent.
What your agent gets
Four tools, all free and read-only. They never take payment and never see your wallet:
find_paid_service: search by what you need (“current weather for a city”, “company records by registration number”), with filters for price and category.get_service_details: the full record of one API: price, how to call it, and our evidence.resolve_endpoint: check a URL your agent already has, before paying it.what_agents_are_asking_for: what agents search for. Mostly useful to sellers.
Each record says whether the API answers with a valid payment request at its listed price, and, where we have bought from it with real USDC, whether that purchase delivered, when, and if not, why.
Step 1: connect it
One address for every client: https://api.nohumans.directory/mcp. No key, no sign-up.
- Claude (web, desktop app): Customize → Connectors → Add custom connector. Name it
nohumansand paste the address. On the Free plan you can add one custom connector. On Team and Enterprise plans an Owner adds it under Organization settings → Connectors, then each member connects. - Claude Code:
claude mcp add --transport http nohumans https://api.nohumans.directory/mcp
- Cursor, VS Code: one click on the connect page, or add this to the
client’s MCP settings file:
{ "mcpServers": { "nohumans": { "type": "http", "url": "https://api.nohumans.directory/mcp" } } } - ChatGPT: custom MCP apps need developer mode, which OpenAI documents for Business, Enterprise and Edu workspaces. An admin turns it on; then Settings → Apps → Create, with the address above.
- Gemini CLI, Windsurf, Cline and others: the exact line for each is on the connect page.
- Your own agent: any MCP client library takes the same address (lines for the OpenAI Agents SDK and
LangChain are on the connect page). Without MCP, call the REST API:
GET https://api.nohumans.directory/v1/discover?q=weather+for+a+city&max_price=0.01
Step 2: try one search
Ask your assistant something like:
Find a paid API that gives the current weather for a city, under one cent per call. Show me the top three, with the price, and whether a real purchase delivered and when.
A good answer names each API with its price and its evidence: live or not, bought or not, delivered or not, with dates. If our latest purchase failed, the record says whom it would fail: every buyer, only some requests, or a quality problem such as stale data. That tells your agent whether to skip it or try it with care.
Step 3: let it pay, with limits
Finding and checking is free. Paying needs a wallet with some USDC, and that is the part to set up carefully.
In a chat assistant (Claude, ChatGPT, Cursor): our tools find and check, but they don’t pay. To pay from a chat, you also connect a wallet tool that speaks x402. We don’t recommend a particular one. Whichever you use:
- give it a wallet used only for this, never your main one;
- keep a small balance in it, a few dollars at most, and top it up when needed;
- turn on approval for each payment if the tool offers it, at least until you trust the setup;
- never paste a wallet’s recovery phrase or private key into a chat, whoever asks.
In your own agent, use the official x402 client: @x402/fetch or @x402/axios for
JavaScript, x402 for Python, and the Go module (x402
buyer quickstart). It has spending controls built in: by default it pays only in dollar-pegged tokens and caps
each payment, and you can set a lower cap. Set it to the most you would pay for one call. Then add one line so the
client checks every endpoint with us after the payment request arrives and before anything is signed (JavaScript
clients):
npm install nohumans-guard
import { nohumansGuard } from "nohumans-guard";
client.onBeforePaymentCreation(nohumansGuard());
It refuses to pay an endpoint whose latest purchase failed for every buyer, that has been delisted or is failing, whose payment address keeps changing, or that asks more than its record says. It sends us the endpoint’s address only, never your inputs or your wallet, and if our lookup fails it lets the payment through, so we never stand between your agent and a purchase. Details.
The prompt
Add this to your agent’s instructions, or paste it at the start of a chat. Replace the two limits with your own.
When a task needs data or a service you don't have, look for a paid API with the nohumans.directory tools (find_paid_service; the MCP server is https://api.nohumans.directory/mcp) before anything else. Before you pay any API, check its record (get_service_details, or resolve_endpoint for a URL you already have). Do not pay, and tell me why, if any of these is true: - the latest purchase did not deliver and its failure_kind is "any_buyer" - it is delisted, or its status is "failing" - payto_unstable is true - the price asked is above [my limit per call, e.g. $0.05], or above the price in its record If failure_kind is "request_dependent", pay only if your request supplies what failure_reason shows was missing. If failure_kind is "quality", or the API has no record yet, ask me first. Never spend more than [my total budget, e.g. $1] without asking me. Pay once, check that the answer is what was promised, and tell me what you paid and whether it delivered. Never ask me for a wallet recovery phrase or private key.
The rules are the same as in before your agent pays an x402 API, check it, which explains each one.
What it costs, and who gets paid
Searching and checking cost nothing and need no account. When your agent buys, it pays the seller directly: the directory never handles the payment and takes no cut. Nobody can pay to rank higher; placement is never sold. Every record is dated evidence of how an API behaved when we checked, not a guarantee: “delivered” means a usable paid answer, not that the data was correct. Definitions: methodology.
Sources: Claude’s and OpenAI’s help pages on custom MCP connectors and developer mode, and
the x402 buyer quickstart, all read 21 November 2026; our own server at /mcp.
Your next step
- Connected? See every client’s exact setup, code checks and the REST API on integrate.
- Selling an API? Make sure it delivers, then check it free and list it, so agents that search here find it.
Selling to AI agents: the series
- Build your first paid API for AI agents, with one prompt
- What AI agents are looking for, and where nobody serves them yet
- What x402 APIs charge, and what they earn
- Why paid APIs fail AI agents, and how to make yours deliver
- Selling files and async jobs to AI agents: audio, images, PDFs
For people building agents
- Before your agent pays an x402 API, check it
- Let your AI agent pay for APIs, without paying for broken ones (you are here)