How Flashfood put its whole company on AI without making everyone an agent expert
Flashfood is a grocery marketplace, not an AI company. It put its whole team on MintMCP, where anyone connects one bundle and starts working. We talked with Yixin Zhu, who leads product and technology.
Getting AI to work meant setting up your own environment
Flashfood is a grocery marketplace. Retailers sell their surplus fresh groceries, shoppers buy them at around half off, and less food goes to landfill. The people doing the work partner with real brick-and-mortar stores and move real food. Most of them want technology to get out of the way, not to become a second job.
So when Flashfood started adopting AI internally, the setup fell on each person. Everyone had to stand up their own harness, credentials, and connectors. Getting Claude to talk to Snowflake meant standing up three or four private MCPs with individual credentials just so a handful of people could prototype. Even the technical group hit a wall.
“We asked someone to share their config file when we were debugging. They posted it in Slack, and it had all their API keys in it.”
Yixin Zhu
Head of Product & Technology, Flashfood
Each new person meant repeating the whole thing, and no one could see how a prompt turned into tool calls or where it broke down. Flashfood needed a layer that did the setup once and put guardrails behind it, so the rest of the company could just use AI.
What Flashfood deployed
Flashfood made MintMCP the layer everyone connects through. Onboarding used to mean figuring out which connectors you needed and whether each one wanted a service account, an OAuth flow, or an API key. Now a new person downloads Cowork, connects to one bundle, and starts working. The whole guide is about four pages of screenshots.
“Download Cowork, connect to one bundle, and you're off to the races. You're onboarded in less than five minutes.”
Underneath, every call now lands in Mint. Where the harness's own analytics only showed usage per model, Yixin's team can see whether a call came from a service account, an agent, or a person, and debug it without asking anyone to send over an error. They can even tell whether a hundred-dollar day was setup, an infinite loop, or someone generating memes.
“We have a bot named Koji, after a Japanese fermentation that gets better over time. It combs the usage, builds skills, and edits the agents' files. The system just gets better over time.”
With the plumbing handled, teams started building on top of it. Skills and agents live in one place, and departments began asking what agents they should build and what skills to release, instead of fixing their own tooling. Two agents now carry most of the usage: one for insightful data, one for real-time knowledge like what's shipping next week or whether the roadmap is on time.
The rollout reached the entire company, and everyone connects through Mint by design. For a team whose real work is feeding the planet, that was the point: buy the layer that works, and keep the engineers on the product.
What Flashfood told us
- RolloutYou're not an AI-native company. How did you end up looking for an MCP gateway?
- “We didn't start from an AI feature. We're trying to solve a real-world problem: partnering with brick-and-mortar retailers to move real food to families. AI isn't at the center of that, but digital transformation is. As we started adopting AI internally to help our staff serve partners and shoppers, it became clear this was still a pretty technical problem, and that's how we stumbled onto the need for Mint.”
- RolloutWhat did the early days look like before a gateway?
- “Everyone had to set up their own environment: their harness, their credentials, their connectors. At one point we had three or four private MCPs stood up with individual credentials, just so the three or four of us prototyping could access things and play with them. Even though we considered ourselves relatively technical, we still had a hard time getting it to work.”
- SecurityWhat broke as you added people?
- “So much of it was repetitive. Every new person meant here's how you match our environment, here's how you do this. When we were debugging, we asked someone to share their config file, they posted it in Slack, and it had all their API keys in it. It's not a rare mistake, but that's exactly what we were seeing every day.”
- RolloutWhat made it click that a gateway was the answer?
- “I saw a LinkedIn post, honestly a little clickbaity, along the lines of 'if you're running MCP servers locally on every machine, here's why you're doing it wrong.' But it was a real problem I was living with every day. When I looked into virtual MCPs and the idea of an abstraction layer, the engineer in me went, yes, that's what's missing. Everyone's harnesses were going direct to the database, and that doesn't scale.”
- RolloutWas security or the user experience the bigger driver?
- “Both matter, but it started with the user experience. Even if your security is perfect, a store owner or a go-to-market person is never going to provision a Snowflake account and wire up the right databases. So we started from how do we make this accessible to everyone, and abstract away the technical details. Through that abstraction you also have to think about guardrails, because compliance matters.”
- RolloutWhat does onboarding look like now?
- “Download Cowork, connect to one bundle, and you're off to the races. All the questions, what connectors do I need, is this a service account or an OAuth or an API key, all of that is gone. I have a guide that's maybe four pages with screenshots, and that tells you how little there is to it. You're onboarded in less than five minutes.”
- SecurityWhat changed on visibility and logging?
- “Anthropic's analytics API is almost comically high level: usage per model, Cowork versus chat versus code, and that's about it. Now every call is in Mint. If someone has an issue, I don't need them to send me the error. I can see whether it was a service account, an agent, or a direct tool call, and debug it. If someone spends a hundred dollars in a day, we can tell whether that was setup, an infinite loop, or just generating memes.”
- ImpactYou built a bot on top of the logs?
- “We have a bot named Koji, after a Japanese fermentation that gets better over time. It combs through the feedback and usage and proactively builds skills and plugins, and edits the agents' files. If there are a lot of errors, it goes and fixes them. If a prompt is burning context deducing the same rules every time, that becomes a skill. The system just gets better over time.”
- ImpactWhat's adoption been across the company?
- “On connection, it's everyone. We rolled it out to the entire company, and no one is not going through virtual MCPs on Mint, by design. Two agents dominate: one for insightful data, and one for real-time knowledge, like what's going live next week or whether the roadmap is still on time. More than 70% of usage is those two in one form or another.”
- ImpactWhat would you tell a VP of engineering who hasn't started?
- “Treat it like an engineering decision, a build-versus-buy decision. You could put three or four engineers on building a gateway, wrapping bundles in MCPs, standing up service accounts, and monitoring all of it, but then you have a full-time team doing just that. We're not an AI company. We partner with retailers, put food on shelves at a discount, and feed families. Unless you've got engineers sitting around with nothing to do, go find a solution that works. Mint is one of the best out there.”
Move from a setup on every machine to one bundle to connect
See how MintMCP gets a whole company onto AI without a team to run the gateway. Read more stories