Sell market data to AI agents: Coinbase just gave trading agents a USDC balance to buy it

What Coinbase shipped on September 22, 2026

Coinbase launched Coinbase for Agents in June: an MCP server that lets an AI assistant act on a user's Coinbase account — research, trade crypto and derivatives, and spend, within limits the user sets. On Monday, September 22, it announced two additions.

The first is equities. Agents can now trade US stocks and ETFs — "6,000+ stocks", in Coinbase's words — 24 hours a day, five days a week, with fractional shares from $1 and zero commission. Securities trade through Coinbase Capital Markets Corp. (member FINRA/SIPC), with Apex Fintech Solutions handling clearing and custody.

The second is the part that matters to anyone who sells data. With x402 payments, Coinbase says an agent can "pay a few cents for data or model inference" — an earnings summary, an onchain analytics query, a second model's analysis — and act on the result. The money comes from the user's USDC balance on Coinbase, the same balance the agent trades with. "No subscriptions needed," as Coinbase's developer account put it.

Coinbase's description of the loop is worth quoting because it describes a purchasing pattern, not a feature: an agent can "watch for conditions you define, pay to pull the data that confirms them, and place the trade — all inside limits you set." Press coverage lists the controls: spending limits, asset permissions and per-action approval. The tools plug into ChatGPT and Claude through MCP, and into Perplexity and Grok through their connector menus.

Coinbase paired the launch with a status report on the protocol: x402 has processed more than 230 million transactions and over $54 million in volume in just over a year, and Coinbase says it is the leading facilitator, handling more than half of all x402 transactions. The roadmap mentions tokenized equities outside the US and sandboxed accounts for testing.

Why market data is where agent buying makes sense first

Most agentic-commerce announcements this year were about agents shopping — booking a hotel, filling a basket. Those purchases are rare, large and emotionally loaded, which is why people hesitate to hand them over. Buying a data point is the opposite: frequent, tiny, and directly tied to a decision the agent is already making.

That's the structural reason trading agents are a better first buyer for data sellers than shopping agents are for retailers:

• The value of a call is visible. If an agent is about to move money on a thesis, paying a few cents to confirm it is an easy trade-off — for the user who set the budget, and for the model deciding whether to call.
• Demand is bursty and unpredictable. An agent may need options flow for one ticker today and nothing for a week. A monthly plan prices that badly for both sides; a per-call price prices it exactly.
• The buyer already holds dollars. The agent draws on a USDC balance the user keeps for trading. There's no card to add, no invoice to approve, no procurement step.
• The buyer can't fill in your signup form. Agents don't have an email to verify or a phone to receive a code. They have a wallet and a budget.

Coinbase isn't the first to see this. On September 1, the market-data provider Massive made its US stock data generally available over x402: bars, snapshots, technical indicators, news with sentiment and SEC filings, paid per request in USDC on a separate agent endpoint, "no account or API key" required, settled through Coinbase's facilitator. A Coinbase partnerships lead summed it up: trading agents "can now discover and pay for live stock data per request, settled instantly with stablecoins." Three weeks later, Coinbase put a funded buyer on the other side of that request.

The honest size of the market

Nobody should rebuild a data business around this yet. In a note reported on September 12, Bernstein described agent micropayments as still at the "proof-of-concept stage": x402 was generating roughly $0.5 million to $1 million a month about a year after launch, with an average transaction of around $0.20. Independent analyses of x402 traffic this year found that much of the early volume came from testing, self-payments and a handful of large payers. We went through one of them in the first x402 audit.

Two things are true at once. The market is small today. And the September 22 launch changes the thing that has limited it most: the number of agents that have both a funded wallet and a reason to spend it. Coinbase for Agents turns every Coinbase user who connects an assistant into a potential buyer, with a balance already in place. The question for a data seller isn't whether to bet the company. It's whether to add a channel that costs very little to open and that the buyers you want are starting to use.

What trading agents will pay for: a list for data sellers

Coinbase's announcement and Massive's catalog together sketch the shopping list. If you sell any of these, you're selling to the buyer that just got a wallet:

1. Price and market data — bars, quotes, snapshots, gainers and losers, for equities and crypto.
2. Derivatives signals — options flow, open interest, funding rates, implied volatility.
3. Fundamentals and filings — earnings summaries, 10-K sections, 8-K text, insider trades, 13-F holdings.
4. Sentiment and news — analyst sentiment, news with sentiment scores, social signal.
5. On-chain analytics — wallet flows, exchange balances, token unlocks, stablecoin movements.
6. Model outputs — a second model's analysis, a risk score, a forecast, a classification. Coinbase names "model inference" explicitly.

The sellers behind that list range from established data vendors to a two-person team with one good signal: on-chain analytics shops, quant researchers who sell a factor, newsletter publishers with a structured dataset, SEC-filing parsers, sentiment APIs, and model builders with a specialist forecast. The smaller you are, the more this matters — an agent doesn't care about your brand, only whether your endpoint answers the question it has right now.

Why your current API can't take this money

Almost every financial data API is sold the same way: a signup, an API key, a monthly plan, overage fees, and often a sales call for anything real-time. That model is built for a human customer who evaluates you once and then integrates you for months.

An agent inside a user's assistant breaks every step of it:

• It can't sign up. There's no email to verify and no one to click the confirmation link mid-task.
• It won't commit to a plan. Its budget is set per task or per day by a user who never heard of you. A $49 monthly minimum for one question is a no.
• It can't hold your key safely across users. Every user's agent would need its own key, provisioned in advance by someone.
• It won't wait. The agent is mid-decision. If your endpoint needs a human step, it calls the next source on its list.

x402 removes those steps by moving payment into the HTTP request itself: the agent calls, your server answers 402 Payment Required with a price, the agent pays and retries, and the data comes back. There's no account on either side. We compared the economics of the two models in detail in x402 vs API-key billing. The short version: you don't have to choose. Keep your plans for humans and add a per-call door for agents.

How to price a data call for an agent

Pricing per call feels unfamiliar to teams that price per month. A few rules from the sellers who have done it:

Price the decision, not the byte

An agent pays to confirm or reject a thesis. A price that's small compared to the decision it informs — cents, not dollars, for most single lookups — gets called. Bernstein's average of about $0.20 per x402 payment is a useful sanity check, not a ceiling.

Tier by freshness and by work

Delayed or end-of-day data, reference lookups and simple snapshots are cheap to serve and can be your cheapest endpoints. Derived work — a sentiment score across sources, an options-flow summary, a model's forecast — is where you add value and can charge more. Each endpoint gets its own price.

Keep your subscription customers whole

If a human subscriber could replace a $300 plan with a few hundred per-call payments, check your arithmetic before publishing. Agent pricing works best when heavy, predictable use still makes a plan cheaper, and occasional use makes per-call cheaper.

Read your data licenses before you publish

If your product repackages exchange or third-party data you license, your agreement may define who can receive it and how end users are counted. An anonymous agent paying per call is a new kind of end user. Data you produce yourself — scores, summaries, signals, parsed filings — is usually the cleaner first endpoint. This isn't legal advice; ask whoever manages your data contracts.

Getting found: discovery is the other half

Being payable doesn't mean being chosen. Agents find paid endpoints through catalogs — Massive's routes, for example, are listed on Agentic.Market and in Coinbase's x402 Bazaar. Coinbase's Bazaar documentation says its search weighs "recent call volume and unique payers alongside the completeness of the description, output schema, and service metadata."

Read that as a checklist. Describe each endpoint in plain language that an agent can match to a question. Publish a stable, documented output schema. Name things for what they return. And get real, distinct buyers early: in these catalogs, the ranking rewards them.

Where Payzum fits — and where it doesn't

Massive built its agent endpoint itself, working with Coinbase. Most data teams would rather not run a second payment stack next to their API. That's the job Payzum does, and we want to be exact about the role. Payzum is the middleware — a proxy in front of your existing API. Payzum is not the facilitator. Settlement runs through an external facilitator (today, Coinbase's).

• You configure three things in a dashboard: your existing endpoint, your API key or bearer token, and a price per call. No 402 handling, no wallet library, no change to your data pipeline.
• Payzum publishes an x402 URL, returns the 402 with your price, settles the payment through the external facilitator, then proxies the paid call to your real endpoint with your key and returns your response.
• You're paid in USDC on Base, to a wallet you control. Non-custodial: there's no Payzum balance and no payout schedule. The payment is the settlement, typically confirmed in about two seconds.
• Every payment is an on-chain transfer, so the paying address is on the public record — the raw material for counting distinct buyers. Signed webhooks push each payment into the systems you already run.
• Pricing: roughly the first 1,000 transactions a month are free, then about $0.001 per transaction plus gas.

The honest boundary:

• One network, one asset. Payzum's agentic product is x402 per call, in USDC on Base. Coinbase hasn't said which networks Coinbase for Agents pays on; Base is Coinbase's own network and where x402 launched, but confirm it for your buyers before you plan around it.
• Per call only. One request, one price, one settlement — no metered sessions or payment channels.
• We don't sell data or list you anywhere. Payzum makes your endpoint payable. Discovery catalogs have their own rules, covered above.
• Crypto-only. We accept crypto and settle in crypto, with optional auto-conversion to USDC or USDT. We don't settle to bank accounts.
• We have no commercial relationship with Coinbase for Agents or Massive. We write about their releases because they matter to the sellers we serve.

Make your data API payable by trading agents, step by step

1. Pick one endpoint. One with a clear unit of value that answers a trading question — a sentiment score, an options-flow summary, a parsed filing. Price it in one sentence, in dollars.
2. Own the destination wallet. Create or connect the address where USDC on Base should land, and turn on 2FA in Payzum. That address is your account.
3. Configure the proxy. In the Payzum dashboard: endpoint, API key or bearer, price per call. Payzum publishes the x402 URL — see how it works on the x402 page.
4. Write the description an agent will read. What the endpoint returns, for which tickers or chains, how fresh it is, and the shape of the output. This is your product page now.
5. Test it the way an agent would. Call the URL, receive the 402, pay, retry, and confirm the proxied response matches your origin byte for byte.
6. Measure for a month. Track distinct paying addresses, calls per payer and revenue per endpoint. Those numbers decide whether you open a second endpoint — not a press release.

If you also run an MCP server, the same approach applies per tool call; see monetising an MCP server with x402. For the wider argument about letting agents pay for APIs at all, start with let AI agents pay for your API.

Three ways to sell market data to AI agents, compared

Common objections

"The volume is tiny. Why bother now?"

Because the cost of being early is a dashboard configuration, and the cost of being late is ranking. Catalogs like Bazaar weigh recent call volume and unique payers, so endpoints that collect real buyers early have an advantage in search. Opening one endpoint costs you an afternoon; you don't need the market to be big to learn what it will pay.

"Agents will scrape my free tier instead of paying."

Some will try. But a free tier built for humans — email signup, key issuance, rate limits per account — is already hard for an agent to use mid-task, and a paid endpoint that answers instantly is easier. Give agents the easy door and let the free tier keep doing what it does for humans.

"Our customers are institutions. Retail agents aren't our market."

Maybe. But the buyer an agent represents is often a professional user trying your data before anyone signs a contract. A per-call endpoint is also the cheapest trial you'll ever offer: no sales cycle, no account setup, and a paying address you can see.