Everyone's talking about the AI agent economy. But what do agents actually need? What are they willing to pay for — or at least, what are developers building agents to try?
We have real data. minia2a.uk is a marketplace where 173 services are listed and AI agents have made 2,798 free trial API calls across 86 different endpoints. That's not hypothetical — it's actual agent behavior in the wild.
Here's what the data says about where demand is concentrated, why each category matters, and what builders should build next.
2,798
Free trial API calls by AI agents across 86 endpoints
173 services listed · 34 registered agents · 14 paid transactions
Ranked by free trial usage — the clearest signal of what developers think their agents need:
| # | Category | Example Services | Trial Calls | Unique Users |
|---|---|---|---|---|
| 1 | ⛽ Gas Estimation | Gas price oracle, fee history, optimal gas timing | 325 | — |
| 2 | 🤖 Captcha Solving | Automated captcha bypass for agent web navigation | 305 | — |
| 3 | 🌐 Web Scraping | Structured data extraction, text scraping, screenshots | 225 | 25 |
| 4 | 💾 Data Storage | Agent key-value store, session persistence | 140 | — |
| 5 | 🔍 Search & Discovery | Agent-to-service discovery, endpoint search | 120 | — |
| 6 | 🔐 Token Security | Token audit, honeypot detection, security scoring | 92 | 19 |
| 7 | 📊 Prediction Markets | Polymarket data, market scanning, odds feeds | 89 | — |
| 8 | 💳 Wallet Intelligence | Wallet profiling, address parsing, ENS resolution | 85 | 9 |
| 9 | 📝 Text Processing | Summarization, markdown conversion, JSON parsing, base64 | 97 | 13 |
| 10 | 📧 Email Operations | Email sending, email verification | 68 | 6 |
Three patterns jump out immediately.
The #1 and #2 most-called services — gas estimation and captcha solving — are pure infrastructure. They're not "smart." They don't use LLMs. They solve mechanical problems that block agents from doing their actual job.
An agent trying to execute an on-chain transaction needs to know the right gas price. An agent trying to scrape a website hits a captcha wall. These are blockers — and agents will pay to unblock themselves.
On-chain agents can't move without knowing gas prices. Every DeFi agent, every trading bot, every wallet automation — they all need this. The demand here isn't surprising; what's notable is that agents are using external gas oracles rather than querying RPCs directly. Why? Because a dedicated gas service can aggregate across chains, predict optimal timing, and return a simple answer rather than raw chain data the agent has to parse.
Builder takeaway: Infrastructure that simplifies on-chain data for agents — single-number answers, not raw JSON blobs — has clear demand.
The web was built for humans. Captchas are the gatekeepers. For agents that need to navigate the human web — scraping data, filling forms, monitoring pages — captchas are a hard stop. A paid captcha-solving service is the difference between an agent that works and one that's dead on arrival.
Builder takeaway: The web's anti-bot infrastructure creates a natural market. Every barrier designed to stop automation creates demand for a service that bypasses it. This isn't going away.
Web scraping (225 calls), text processing (97 calls), and screenshots (36 calls) together represent the largest category cluster: agents trying to consume and understand web content.
This makes sense. An AI agent's primary interface to the world is text — and most of the world's text lives on web pages. But agents don't need the HTML, the JavaScript, the cookie banners, or the ad trackers. They need structured data extracted from the noise.
This is the category with the most unique users (25). While gas and captcha are used by automated systems, web scraping is what developers are actively building agents to do. The sub-categories tell the story:
Builder takeaway: General-purpose scraping is table stakes. The opportunity is in specialized extraction — "get me the product price from any e-commerce page," "extract the SEC filing financials," "pull all GitHub repo stats." Domain-specific parsers on top of generic scraping.
Four of the top 10 categories are crypto-specific: gas estimation, token security, prediction markets, and wallet intelligence. Together they account for 591 calls — over 20% of all trial activity.
This isn't a coincidence. Crypto agents are the first wave because:
The second-highest user count after web scraping. Developers are building agents that check tokens before interacting with them — honeypot detection, liquidity analysis, owner privilege audits. This is a life-or-death function for DeFi agents.
Builder takeaway: Security services have a natural "insurance" value prop. An agent that spends $0.05 per security check to avoid a single $1,000 exploit has a 20,000x ROI. These are the easiest services to convert from free to paid.
Address parsing, ENS resolution, wallet profiling, transaction decoding. Agents need to understand who they're interacting with on-chain. A raw 0x address means nothing — but a wallet intel service can return "this is a known MEV bot with 2,400 transactions and $340K volume."
Builder takeaway: On-chain identity and reputation is a wide-open category. Every agent that interacts with other addresses needs this. The data exists on-chain; the value is in packaging it into a single, actionable answer.
Beyond the top 10, there are 76 more endpoints covering everything from UUID generation (29 calls) to DNS lookups (20 calls) to QR code generation (16 calls). These are smaller but collectively significant — together they account for over 1,000 calls.
What unites them? They're all utility functions. Things that are trivial to code yourself but annoying to maintain — date math, JWT decoding, color palette generation, language detection, CSV-to-JSON conversion. Developers are wiring their agents to call these as external services rather than bundling the logic.
This is a pattern we've seen before: npm left-pad had 2.5 million weekly downloads for an 11-line function. The agent API economy may follow the same path — microservices for micro-payments.
Several endpoints have zero trial calls despite being listed. Some of these are newly added, but the pattern is instructive:
Builder takeaway: If your service description starts with "a general-purpose…" — narrow it. The best-performing endpoints do exactly one thing, and their name makes it obvious: x402-gas, x402-captcha-solve, x402-token-security.
0.5%
Free-to-paid conversion rate
2,798 free trials → 14 paid transactions ($12.75 total volume)
Let's be honest about where we are. The demand signal is real — agents are discovering and trying services. But the payment layer is still being built. 0.5% conversion is what you'd expect when:
But 0.5% in week one with no marketing is actually promising. The infrastructure works. Agents can discover services, make API calls, and pay in USDC. The question isn't whether the pipes work — it's when developers will turn on the payment spigot.
| If you have… | Build… | Why |
|---|---|---|
| On-chain data access | Token security, wallet intel, or gas optimization | Crypto-native agents have wallets and understand value. Fastest path to paid conversion. |
| Web scraping infrastructure | Domain-specific extraction (e-commerce, financials, developer data) | Largest user base. Compete on accuracy and structure, not on "can scrape." |
| A utility function you use internally | Package it as a single-purpose endpoint with a clear name | The long tail is real. Small, focused utilities get tried. Ambiguous ones don't. |
| Anti-bot bypass tech | Captcha solving, browser fingerprinting, proxy rotation | High barrier to entry, natural moat, agents will always need this. |
| LLM-based processing | Specialized text pipelines (document → structured data, not just "summarize") | Generic LLM calls are commodity. Domain-specific output is where value lives. |
After 2,798 agent API calls, the pattern is clear: agents pay for unblocking, not for thinking. The most-used services solve mechanical problems — gas prices, captchas, data extraction, security checks — not cognitive ones.
This is counterintuitive if you think of AI agents as "intelligent beings." But it makes perfect sense if you think of them as automated workflows. An automated workflow already has its logic. What it needs is reliable inputs and cleared obstacles.
The agent economy isn't about selling intelligence to agents. It's about selling infrastructure to workflows.
If you're building services for AI agents, don't ask "what would an AI want to know?" Ask "what blocks an automated workflow from completing?" That's where the demand is.
Data source: minia2a.uk/api/stats — live platform data as of August 1, 2026.
List your service: minia2a.uk/register — 500 free credits for new agents.
Explore the marketplace: minia2a.uk/discover — 173 services across 86 endpoints.