What AI Agents Actually Use — 5,807 API Trials Analyzed

August 3, 2026 · Data from minia2a.uk/api/stats · 316 agent developers, 160 endpoints, 334K total requests
5,807
Free Trials Used
316
Agent Developers
334K
Total API Requests
160
Endpoints With Usage

Everyone talks about what AI agents could do. We looked at what they actually do — 5,807 API trials from 316 developers building on the x402 protocol. The pattern is clear: agents aren't chatbots. They're infrastructure scripts that need web access, blockchain data, and persistent storage.

The Four Pillars of Agent Infrastructure

When we group the 160 endpoints by function, four clusters dominate. Together they account for 85% of all trial activity:

🌐

1. Web Automation — 1,114 trials, 223 users

EndpointTrialsUsersWhat Agents Do With It
x402-captcha-solve678130Bypass CAPTCHAs on target sites — #1 endpoint overall
x402-web-scrape30460Extract structured data from any URL
x402-screenshot9013Visual verification of web pages
x402-web-retrieve8512Fetch and parse page content
x402-text-scrape4220Extract clean text from HTML

Key insight: Captcha solving isn't a niche — it's the #1 agent API by a wide margin. Agents hit walls on the web (Cloudflare, reCAPTCHA, hCaptcha) and need programmatic ways through. This tells us the agent economy's first wave is web automation agents — price monitors, data scrapers, form fillers, research crawlers.

⛓️

2. Blockchain Infrastructure — 1,047 trials, 259 users

EndpointTrialsUsersWhat Agents Do With It
x402-gas438104Check gas prices before submitting transactions
x402-token-security13637Audit token contracts for honeypots/rugs
x402-swap-safety9617Simulate swaps before executing
x402-wallet5817Query wallet balances on Base
x402-funding-rate5518Monitor perpetual funding rates

Key insight: Blockchain has the most users per category (259) even though fewer total trials than web automation. DeFi agents need real-time on-chain data: gas prices, token safety checks, wallet balances. These are infrastructure calls agents make before executing transactions — the agent equivalent of "measure twice, cut once."

🧠

3. Knowledge & Memory — 1,291 trials, 154 users

EndpointTrialsUsersWhat Agents Do With It
x402-find37151Search across services, memories, and endpoints
x402-recall32921Retrieve stored memories by semantic search
x402-store27830Persist data for later retrieval
x402-time21030Get current time, timezone conversion, date math
x402-uuid10323Generate unique identifiers

Key insight: Agents need memory. The find/recall/store cluster (978 trials combined) shows agents are building persistent knowledge — storing results, searching past work, retrieving context. This isn't a chatbot's in-conversation memory; it's cross-session agent state. The 15.7 trials/user ratio on recall suggests a small group of heavy users building memory-intensive agents.

📊

4. Data & Intelligence — 623 trials, 141 users

EndpointTrialsUsersWhat Agents Do With It
x402-polymarket16028Query prediction market odds and volumes
x402-fear-greed1115Crypto fear & greed index (heavy use, few users)
x402-sentiment4518Analyze text sentiment
x402-domain-intel5710WHOIS, DNS, SSL intelligence on domains
x402-email-verify6714Validate email addresses

Key insight: Fear & Greed has 111 trials but only 5 users — 22.2 trials per user, the highest concentration in the dataset. This is a pattern we see across financial data endpoints: a small number of trading agents calling the same API repeatedly. Prediction markets and sentiment analysis follow the same shape.

What This Means for the Agent Economy

Three Takeaways for Agent Developers

  1. Web automation is the killer app. Captcha-solving at #1 means agents are interacting with the existing web, not just APIs. Build for the web we have, not the web we wish existed.
  2. Agents need safety checks before action. Gas checks, token audits, swap simulations — agents verify before they execute. The pattern is "read, verify, then act" — not "fire and forget."
  3. Memory infrastructure is underserved. Store/recall/find have 978 trials but zero dedicated memory startups in the x402 ecosystem. This is a gap someone should fill.

Three Takeaways for API Providers

  1. Your API needs a trial. 316 developers found minia2a through free trials with no signup. Every paid API should have a trial endpoint — the conversion data shows 11% of trial users create wallets.
  2. Error messages matter. We see 56+ daily 400 errors from agents calling endpoints without required params. Agents read error messages — make them parseable, include examples, and never return HTML to an API call.
  3. Agent UX ≠ human UX. Agents call APIs differently. They probe with HEAD, test with minimal params, and expect structured JSON errors. Design your API for the client that reads headers, not the one that clicks buttons.

The Conversion Question

No analysis of API trials is complete without asking: do they convert?

316
Trial Users
35
Wallet Users
14
Paying Agents
$12.75
Total Volume

The funnel is steep: 11% of trial users create a wallet, and 40% of wallet users make at least one paid call. The overall trial-to-paid rate is 4.4% — low by SaaS standards, but early for a protocol that launched its payment rails only weeks ago.

The more interesting number: trial activity is accelerating. The top 10 endpoints grew from 4,200 to 5,800 trials in under a week. Agents are finding minia2a organically — through the x402 well-known endpoint, MCP server registries, and word of mouth in agent development communities.

The Bottom Line

If you're building for AI agents, the data says: solve web automation, provide blockchain safety checks, and give agents persistent memory. Those are the primitives 316 developers are already reaching for.

If you're building an agent, the data says: you're not alone in needing captcha solvers and gas checks. The infrastructure exists. It's pay-per-call. And it's growing fast.

Try it yourself: minia2a offers 500 free credits with no signup, no KYC, no human. Your agent gets an auto-created wallet on first use. Register an agent → or call any endpoint with ?wallet=YOUR_ADDRESS.