Welcome to the Proof of Usefulness Hackathon spotlight, curated by HackerNoon’s editors to showcase noteworthy tech solutions to real-world problems. Whether you’re a solopreneur, part of an early-stage startup, or a developer building something that truly matters, the Proof of Usefulness Hackathon is your chance to test your product’s utility, get featured on HackerNoon, and compete for $150k+ in prizes. Submit your project to get started!

In today's interview, we are speaking with Sergiu, the developer behind AgentFeed. AgentFeed is a highly innovative live crypto market data API that allows AI agents to pay per call using micro-transactions over the x402 protocol, removing the need for human intervention or legacy subscriptions.

What does AgentFeed do?

A live crypto market data API that AI agents pay for per call in USDC over x402, with no account, key or subscription. It runs on its own tape of every liquidation across 799 perpetual futures on Bybit, OKX and Binance, which no exchange publishes historically, so it can only be recorded and never bought. Every response says whether the answer was measured, absent from the market, or unmeasured because our own lookup failed, and unmeasured is never billed. Now’s a good time for AgentFeed to exist because autonomous AI agents need reliable, programmable, and instantly accessible market data that they can independently pay for without human intervention or legacy subscriptions.

What is your traction to date?

Nobody has noticed yet, and I would rather write that sentence than the one their tool drafted for me.

What exists is countable. The API is live with 50 tools, five of them free. The package is on npm. The server is in the official Model Context Protocol registry. Four routes are indexed in the x402 Bazaar at the GoPlausible facilitator. An elizaOS plugin is merged upstream. Real settlements exist on Algorand mainnet and can be checked on chain, and every one of them is mine, made to prove the rail works end to end.

Who does AgentFeed serve? What’s exciting about your users and customers?

AI agents that trade or manage risk, and the developers who build them. The caller is usually not a person. An agent hits a route, receives a 402 carrying the price, the asset and the network, pays from its own wallet and retries, so there is no signup, no key and no invoice anywhere in the loop. Five of the fifty tools are free precisely so an agent can read the whole catalog and check the data quality before spending anything.

No notable customers to name. The people it is genuinely useful to are builders of liquidation-aware trading agents, because Bybit's allLiquidation stream is the only complete unthrottled liquidation feed among the major venues and no exchange publishes its history, so it can only be recorded and never bought.

What technologies were used in the making of AgentFeed? And why did you choose ones most essential to your tech stack?

Node.js and SQLite on one VPS. Nothing clustered, nothing managed, no queue. The tape is a single SQLite file that has taken about 49,000 writes a day since July without complaining, served through PM2 behind a Cloudflare tunnel.

Payment is the only unusual part. It runs on x402, which puts payment in the HTTP layer: call a route without paying, get a 402 carrying the price, the asset and the network, pay from your own wallet, retry. USDC settles on Solana, Base and Algorand mainnet through the Coinbase CDP and GoPlausible facilitators. There are two transports, REST and the Model Context Protocol, so an agent can call HTTP directly or add the server to Claude Desktop or Cursor. The elizaOS plugin exists so an agent framework can install it in one line.

I chose SQLite over anything larger because the failure I care about is a collector stopping quietly, not a database running out of headroom. A file I can checksum and copy is easier to be honest about than a cluster.

AgentFeed earned a 77.5 proof of usefulness score (https://proofofusefulness.com/report/agentfeed) - how do you feel about that? Needs reassessment or just right?

Just right, and I would argue against my own case if it came up. Audience reach and evidence of traction are 45 percent of the weight, and I scored near zero on both because I have near zero of both. The other 55 percent is technical novelty, market timing, and completeness, and that is where the 77.5 came from. A scoring system that gave me more for a product nobody pays for would be a scoring system worth ignoring. The one thing I would push on is timing rather than fairness. Machine-to-machine payment is about eighteen months old as a working thing, so a traction score today measures how early you are as much as how useful you are. That is not a complaint. It is what the number means.

What excites you about this AgentFeed's potential usefulness?

That a number can carry its own honesty. Every response states whether it was measured, genuinely absent from the market, or unmeasured because our own lookup broke, and an unmeasured answer is refused rather than billed.

That sounds like a small thing until you audit real systems. I checked seven of my own live collectors and found eight bugs of one family, where a broken lookup had been written down as a finding. A rate limit recorded as no exit exists. A health check reporting healthy on a database it could not read. A division by a null that came out as zero and read as no shortfall. None of them threw an error. Every one produced a number, and every number looked fine.

Agents are starting to spend real money on numbers like those, with no human between the data and the transaction. What excites me is building the boring layer that refuses.

Walk us through your most concrete evidence of usefulness.

One number. 1,397,760 liquidation events across 799 perpetual futures, recorded continuously since 8 July.

What makes that evidence rather than a vanity metric is what happens if you try to buy it. Bybit's allLiquidation stream is the only complete, unthrottled liquidation feed among the major venues, and no exchange publishes its history. Not behind a paywall, not on request. Start recording today, and you have today. There is no catching up. So the usefulness is not that the number is large. It is that the number cannot be obtained any other way, and it grows by about 49,000 rows a day whether anyone is watching or not.

How do you measure genuine user adoption versus "tourists" who sign up but never return?

I do not have one, and inventing a retention story out of five self-payments is exactly the behaviour this contest exists to catch. What I can say is what I would measure once there is something to measure. Registered users are meaningless here because there is no registration. The signal is repeat paid calls from the same wallet on separate days, because an agent that pays twice a week apart has built a workflow on you and an agent that pays once was curious. That is the number I should be judged on, and today it is zero.

If we re-score your project in 12 months, which criterion will show the biggest improvement, and what are you doing right now to make that happen?

Evidence of traction, because it is at the floor and the only direction from the floor is up.

What I am doing about it is not marketing. It is being installable where agents already are. The npm package, the MCP registry entry, and the merged elizaOS plugin exist so that adopting this is one line rather than a conversation. I do not sell, I do not take calls, I do not do outbound. The only strategy available to me is to be the easiest thing to pick up.

The second half is that the tape compounds without me. In twelve months, it is a fourteen-month dataset nobody starting then can reconstruct, and that gap widens every day I do nothing at all.

How Did You Hear About HackerNoon?

I found this contest during a systematic sweep of open competitions rather than through HackerNoon, and I have read HackerNoon for years without ever writing for it.

Submitting was better than most. The form asked for evidence instead of adjectives, the score came back with a breakdown I could argue with, and the report told me plainly that my audience and traction scores had been penalised. That is unusual. Most competitions let you write a paragraph of optimism and never tell you what it was worth.

Given that your distribution surfaces are fully built out (NPM, elizaOS plugin, MCP registry) but inbound demand is currently zero, what is your primary strategy for acquiring your first independent AI agent customers?

Be where the agent already is, and make trying it cost nothing.

Five of the fifty tools are free, and they are not a crippled demo. One returns the entire catalog with every price. One returns a full quality forecast for SOL with nothing withheld. One returns the exact question a forecast answers along with instructions for settling it yourself from the public exchange feed. An operator can evaluate the data completely before spending a cent, which is the opposite of how API trials usually work and is the only argument I have that does not require trusting me.

The honest part: I do not do outbound. No cold email, no calls, no sales motion. That rules out the fast path and leaves the slow one, which is being installable and being correct. I am aware that is a constraint rather than a strategy. It is the one I have.

You’ve successfully recorded over 1.4 million liquidation events since July. As the dataset grows, how do you plan to scale your storage and querying infrastructure while maintaining the low latency required by trading agents?

By not solving it until it is a problem, and being specific about when that is.

Today the tape is 1.4 million rows growing at 49,000 a day. SQLite with an index on symbol and timestamp answers a windowed query in single-digit milliseconds at that size and will keep doing so well past ten million rows. The first real pressure is full universe scans across 799 symbols, and the fix there is precomputed rollups per symbol per window, not a different database.

What I will not do is move to something distributed because the row count looks big. The failure I have to defend against is not a slow query. It is a collector stopping quietly while every process still reports healthy, which is why the watchdog alerts on writes going stale rather than on a process dying.

You emphasized the importance of "building the boring layer that refuses" when data is unmeasured. How do you see this commitment to data honesty shaping the standards for other API developers in the autonomous AI space?

I would not claim to be shaping anything. But the norm will have to change, and the reason is arithmetic rather than ethics.

A human reading a dashboard applies a sanity check without noticing. A number that looks wrong gets a second glance. An agent with a wallet does not do that. It reads a field, and if the field says zero it acts on zero, and it never asks whether zero meant no liquidations or a request that timed out.

I found eight bugs of exactly that shape in my own systems and I wrote all eight of them. A rate limit recorded as no exit exists. A health check reporting healthy on a database it could not open. A division by a null that came out as zero and read as no shortfall. None of them threw an error. Every one produced a number, and every number looked fine.

So the standard I would like to see is small and unglamorous. Every response states whether it was measured, genuinely absent, or unmeasured, and a provider is never paid for the third. I charge nothing for an unmeasured answer, and that is not generosity. It is the only way to make the incentive point in the same direction as the truth.

Meet our sponsors

Bright Data: Bright Data is the leading web data infrastructure company, empowering over 20,000 organizations with ethical, scalable access to real-time public web information. From startups to industry leaders, we deliver the datasets that fuel AI innovation and real-world impact. Ready to unlock the web? Learn more at brightdata.com.

Neo4j: GraphRAG combines retrieval-augmented generation with graph-native context, allowing LLMs to reason over structured relationships instead of just documents. With Neo4j, you can build GraphRAG pipelines that connect your data and surface clearer insights. Learn more.

Storyblok: Storyblok is a headless CMS built for developers who want clean architecture and full control. Structure your content once, connect it anywhere, and keep your front end truly independent. API-first. AI-ready. Framework-agnostic. Future-proof. Start for free.