
Suncly Tested 554 AI Agents. Only 45% Did What They Promised.
Founder Stories by Built in Baltics. Mihkel Hõbemägi in conversation with Kasper Muutra, co-founder & CEO of Suncly (with Kristjan Hertsmann).
In this article
Kasper Muutra and Kristjan Hertsmann tested 554 public AI agents against their own published descriptions. 97% passed the basic checks. Only 45% actually did what they claimed.
The checks everyone runs pass almost everything, while the claims buyers rely on hold true less than half the time. The two TalTech students founded Suncly on 4 October 2026 to close that gap: they test AI agents and MCP servers the way a real user would, then sign the evidence. Six weeks in, they have paying customers, a self-serve console and a price that has gone up four times.
How It Started
The idea came from a first-year networking class at TalTech. Studying the OSI model, the stack of layers that lets any two computers talk to each other, they noticed AI agents have no such stack. And no layer for trust at all.
So they wrote a safety and security schema for A2A, Google's new protocol for agents talking to other agents. One of the Google engineers behind the protocol reached out with feedback. That was the push.
"Agents tell the world what they can do, and nobody checks."
Suncly exists to be the ones who check. Kasper is the CEO, Kristjan the CTO.
Building From the Baltics
Estonia shaped the core of the product. Estonians sign contracts, tax returns and even votes digitally, so Kasper and Kristjan grew up believing a claim only counts once someone signs it. Suncly applies that to agents: it doesn't just test an agent, it signs the evidence.
The flip side is that nobody in Estonia buys what Suncly sells. Their first sales email went to the US, and their sales calls start at 5 pm.
How It Works
A team adds its MCP server or AI agent to the Suncly console and sees the price before anything runs. Suncly then calls every tool the way a real user would, on the team's test account. The result is a signed report: every claim marked kept, broken or untested, with evidence, tied to the exact version tested.
The typical customer is a small team about to submit to the ChatGPT or Claude app directory, where a failed review costs weeks and often comes with no clear reason. With a Suncly report, the guessing disappears. They know exactly which tool to fix.
Under the hood, a Python engine runs each test in its own isolated container on AWS in Frankfurt and treats the agent under test as hostile: whatever it returns is data, never an instruction. Deterministic checks decide what they can. An AI judge, asked three times, handles the rest. Every report is cryptographically signed.
The first proof came from Moviegroupies. They fixed what Suncly's test found and were approved into OpenAI's app directory the next day.
"That's when we knew."
Since then, a customer has come back to book a second test, and the price per test has gone from $100 to $400 while customers kept saying yes.
The Hard Part
Two weeks in, on 20 October, the only team that had promised to pay pulled out. No customers, no revenue.
"We honestly asked ourselves whether we had built something nobody wanted."
They gave themselves one sprint to find a single team that would pay, and wrote only to teams about to submit an MCP server to an app directory. The next day, Moviegroupies paid.
The same day, an investor passed. The day after, Suncly's Reddit, GitHub and Discord accounts were shadowbanned, cutting off the channels they had used to reach builders. They moved everything to email. Within a week of that first sale, they had three paying customers.
They had expected customers to be building A2A agents, because that's where Suncly started. The teams with urgency and a budget were building MCP servers stuck in directory review. Finding that out took their first two weeks.
They also assumed customers wanted Suncly to do the testing for them. But the first reports took six hours of manual work each, and customers said they would rather run a test themselves, at any hour. So within five weeks of founding, Suncly became a self-serve console.
"Let the customer design the product."
What's Next
Kasper and Kristjan point to 1909, when John Moody started rating railroad bonds. Ratings became how investors decided what to trust. They think AI agents are at that moment now. By their count, more than 700 new MCP servers appear every day, and companies plug them into email, data and payments unchecked.
The plan for the next year: the self-serve console, a monthly benchmark on users' real tasks, the first enterprise contracts, a second engineer, and sales run from San Francisco. In three years, they want a Suncly rating to be what a directory, an enterprise or an insurer checks before an agent is allowed to connect, delegate or pay.
To get there, they're raising a €350k pre-seed.
The Toolbox, and One Book
The tool they can't live without is Claude Code. Two founders, no other engineers, and a self-serve console in five weeks. It works like a disciplined third teammate because they give it strict written rules, such as never seeing a real secret and writing a failing test before every fix.
The book is Zero to One by Peter Thiel, which the two borrowed together. Beyond that, they listen to what the big names in AI are saying: Dario Amodei's and Sam Altman's podcasts and interviews. For inspiration they look to Alexandr Wang (Scale AI) and Patrick Collison (Stripe). Every week they also scrape the web for incidents caused by AI agents and cover them on Suncly's site.
And one funny coincidence: a TalTech lecturer assigned them Princeton's "AI Agents That Matter" paper as homework. So at school, they're reading about the exact problem they're building a company around.
At a Glance
| Company | Suncly |
|---|---|
| Founders | Kasper Muutra (CEO), Kristjan Hertsmann (CTO) |
| Founded | 4 October 2026 |
| Based in | Estonia |
| What it does | Tests AI agents and MCP servers against their own claims and signs the evidence |
| Ideal customer | Small teams about to submit an MCP server or agent to the ChatGPT or Claude directory |
| Research | 554 public agents tested: 97% passed basic checks, 45% did what they claimed |
| Traction | 3 paying customers in the first week of sales, one repeat booking, price per test up from $100 to $400 |
| Stack | Python, isolated containers on AWS (Frankfurt), deterministic checks plus an AI judge, signed reports |
| Raising | €350k pre-seed |
| Website | suncly.com |
| Suncly on LinkedIn |
How You Can Help
- Shipping an MCP server or agent? Test it with Suncly before you submit to the ChatGPT or Claude directory.
- Running agents inside your company? Suncly wants a few design partners for its monthly benchmark.
- Angels: they're raising a €350k pre-seed and would love Baltic founders on the cap table.
- San Francisco: introductions, and an engineer who knows MCP or A2A deeply.
Write to team@suncly.com or try it at suncly.com.
I talk to Baltic founders about what they're building and what broke along the way. Building something? Message me on LinkedIn.
Written by
Mihkel Hõbemägi · Co-founder, Built in Baltics