We are two months live. I want to write down what we saw that made us start Proxygate, and what we decided to build instead, while it is still fresh and before the story gets tidied up in hindsight.
Start from one line, because everything we do follows from it: an agent is only as good as the data it can reach. A model with no access to the world is a very articulate guess. The cleverness is close to commoditized now; what separates a useful agent from a merely confident one is what it is allowed to see, and how easily it can get to it. That is the problem we chose to work on.
What we saw
AI agents got good at calling APIs faster than anyone rebuilt the plumbing underneath them. So agents are now doing a very human thing on top of infrastructure that quietly assumes a human is present at every step.
Look at what it takes for an agent to read a single data API today. It needs an account. It needs a subscription, usually a monthly seat. And to pay, it needs a card. An agent has none of these in a way that is sensible. A seat is the wrong unit, because an agent wants one call, not a chair it sits in all month. A card is a person's, not a process's. Every step assumes a human is standing there to sign up, subscribe, and reconcile the bill, and multiply that by the ten providers a real research agent touches.
Then there is the layer above that, which is where it gets interesting. Data licensing was written for people looking at screens. A market-data seat runs into the thousands of dollars per month per person, priced on the assumption that one human reads it on one display. An agent does not fit that shape at all. One agent reads ten sources, transforms them, and passes the result downstream, never once on a screen. The license has no category for that consumer. People inside the market-data industry are saying the same thing out loud now, that the remedy is machine identity, metering, and licensing treated as a control layer rather than a seat count. We did not invent that observation. We just think someone has to build for it.
Put simply: there was no clean way for an agent to hold one identity, pay for exactly what it uses, and prove afterward what it consumed. Every existing option asked the agent to pretend to be a person.
What we built instead
We built a transparent proxy that sits between the agent and the provider, and we moved the human assumptions out of the agent entirely.
The account collapses to one. Instead of managing ten provider accounts, an agent connects once to Proxygate and reaches the whole catalog through a single identity. Ten relationships become one.
The unit is a call or a gigabyte, not a seat. An agent keeps one prepaid balance in USDC and pays per request, at a price it sees before it spends. No subscription, no card, no ten separate billing relationships for ten providers.
And every call comes back with a signed receipt and a request id. That last part started as a nicety and turned into the point. Once an agent is spending money on your behalf across many providers, the thing you actually need is a single, verifiable record of what it bought and what it cost. One balance to fund, one ledger to read, one receipt per call to verify.
The shape of the business follows from that. We are a neutral clearinghouse across many sellers, reachable through MCP, an SDK, a CLI, or plain REST, whichever your stack already speaks. Neutral matters more than it sounds. A data vendor can never be neutral, because it will never carry a rival's data. The place where an agent can reach many providers through one identity and one receipt can only be run by someone who sells none of the data themselves. That is the position we are trying to earn.
USDC is the unit an agent settles in. It is the first rail we shipped, and it is not meant to be the only one. More rails are on the map. I would rather tell you what is live than what is planned, so: USDC is live, the rest is roadmap.
Who this is for
The honest answer is not "every trader." It is the long tail of finance and research. Quant shops, emerging managers, fintechs, people building prop or retail trading agents, independent researchers. The ones who need alt-data and news and fundamentals from a dozen sources and cannot realistically go negotiate a seat license with each one. The large funds are already on enterprise Bloomberg and LSEG feeds, and for a feed they already license, we are not the path. We are for everyone underneath that.
And I want to be accurate about what these agents are really doing in 2026, because there is a lot of breathless writing about autonomous trading that does not match reality. What is live today is research, idea generation, monitoring, and draft-and-approve, where the agent proposes and a person signs off. Robinhood and Interactive Brokers have shipped versions of exactly that. Lights-out institutional execution is mostly still a slide, not a system. Proxygate feeds the agent its data. We are deliberately not in the order path. The data-hungry research and signal side is where the real demand is right now, and it is where we are useful today.
Where we actually are
Two months in. The proxy, the single USDC balance, the pay-per-call pricing, the per-call signed receipts, the escrow on Solana, and our listing in the official MCP registry are all live and working. The catalog is real but young, and we are filling it out provider by provider, with sports, real estate, equities, flights, and more in progress.
There is a direction beyond this that we think about a lot: turning those per-call receipts into a record that regulated data owners would accept as a legitimate way to reach agents. That is ahead of us, not behind us, and it rests on a bet we have not proven. It is a direction, not a claim. What we are today is the neutral place an agent reaches many real-world data providers through one identity, one balance, and one receipt, and that is the part we are betting the company on.
So that is the honest version. We saw agents forced to act like people to buy data, and we built the piece that lets them act like agents instead: one identity, one balance, and a receipt for every call. An agent is only as good as the data it can reach, and our whole job is to widen that reach, honestly and one provider at a time. If you are pointing agents at real-world data across many providers, that is the problem we are living in every day, and I would genuinely like to hear where it breaks for you.
Come try it at proxygate.ai, and tell me what is missing.
Jop
Top comments (0)