Coursera rolled out governed AI to 2,000 employees without slowing developers down
Coursera reached 100% AI adoption, then needed governance to keep it secure. We spoke with their CTO and engineers on securing MCP connections, giving teams approved tool sets, and scaling AI beyond engineering.
Adoption was at 100%. Enablement wasn't.
Coursera went all in on AI, and adoption reached 100%. Engineers tried every new coding tool, use of MCP grew as open-source servers appeared, and teams wanted to connect company data to the AI tools they used.
But no one could say exactly what was connected or where. API tokens were scattered across the company. Developers with elevated permissions were “just letting LLMs run wild,” as Senior Software Engineer Mark Arslan put it. They were using unvetted MCP servers and skills from the internet and pasting credentials directly into models. If someone asked which MCP tools Coursera officially supported, there was no clear answer.
“My concern was less about the technology itself and more about the lack of visibility and consistency across the company.”
Mustafa Furniturewala
CTO, Coursera
The wake-up call came from outside: the industry saw popular open-source packages compromised, with malicious code quietly using the AI tools already installed on developer machines to hunt for secrets. Incidents like these made the risk concrete. The lesson was clear: these tools needed a controlled boundary, and Coursera needed visibility into what crossed it.
What Coursera deployed
Coursera evaluated MCP gateways when the protocol was only months old. MintMCP stood out for how little it required Coursera to change. For AI coding tools, it integrated through hooks that expose tool activity for validation, instead of routing all traffic through the gateway. Secret detection and governance controls were built in, rather than added through a third-party classifier. MintMCP could also host Coursera's own MCP servers, so internal tools did not require a separate system.
“I didn't want someone to have to add 60 different MCPs. If they're a developer, they can just install a developer package and they're ready to go. If they're in product management, they can install a productivity package and they're ready to go.”
Mustafa Furniturewala
CTO, Coursera
Coursera manages access through tool bundles. MintMCP's virtual MCPs— controlled bundles of servers and tools—let the team enable or disable individual actions. An agent, for example, can read a repository without being allowed to fork it or change its permissions. Coursera created one developer bundle, starting with Atlassian and GitHub, then adding Sourcegraph and Sumo Logic. Developers connect once and automatically receive each new tool the team adds, with no new rollout.
“It's transparent to the end user. Even if they're not using MCPs themselves, we transparently added the hooks and could put governance around how people use their agents.”
Mark Arslan
Senior Software Engineer, Coursera
The model then expanded beyond engineering. Around 300 engineers are effectively fully covered, and other teams now have bundles for the tools they need: Figma for designers, product tools for product managers, and Salesforce for business development. Coursera has rolled the system out to 2,000 people in total, and more than a quarter of its AI usage now comes from non-technical teams.
What Coursera told us
- SecurityWhat was happening with AI at Coursera before MintMCP?
- “AI adoption was accelerating rapidly across the company, but the supporting infrastructure and governance were still catching up. MCP usage was growing quickly as teams looked to connect internal data with AI tools, often in decentralized ways. My concern was less about the technology itself and more about the lack of visibility and consistency. If someone had asked, ‘What MCP tools are available across Coursera?’ there wasn’t a single, reliable answer.”Mustafa Furniturewala, CTO
- SecurityWas there a specific wake-up call?
- “We've all seen popular open-source packages get compromised, where someone uploads a version that does things it's not supposed to on developer machines. It was eye-opening to see what could happen if you didn't control the secrets around MCPs or tokens. That was one of the factors why I started looking at MCP gateway solutions.”Mustafa Furniturewala, CTO
- EnablementWhy not just block the tools?
- “The security concerns are completely valid; that's why you need some governance, some structure to using these tools. But when you block certain tools, you see people use patterns that are unsafe. If you block MCP access, they'll start copying data from one place to another or creating some temporary data store, which is worse from a security perspective. The primary goal is to enable everyone, so my first instinct is not to block access but to see what is the right way to enable people with the tools.”Mustafa Furniturewala, CTO
- RolloutWhat did you need from an MCP gateway?
- “We did an analysis of MCP gateways and found MintMCP to be easy to set up and use. We needed to be able to host our own MCPs, and we didn't want a different solution for that, because it needs to be done in a secure way as part of the gateway setup. The other thing that's important is setup time. I didn't want someone to have to add 60 different MCPs. That enablement needs to either be automatic or a very simple click somewhere.”Mustafa Furniturewala, CTO
- RolloutHow do you control which tools people can use?
- “Virtual MCPs. We bundle multiple servers and control individual tools. We knew LLM users shouldn't be able to fork a repo or change repo permissions; we want that done manually if at all. Being able to click tools on and off, and roll out new MCPs to a bundle, means a developer points at the dev bundle once and automatically gets whatever we push out next. We don't have to send anything around.”Mark Arslan, Senior Software Engineer
- SecurityHow does the security team see it?
- “We work hand in hand with them. They wanted to know what people were doing and to make sure usage wasn't creating unnecessary risk. The insights dashboard shows the servers people use and how they're using them, and the rules engine (built-in rules, custom tools, and a tool allow-list) lets us enforce it. We blocked all sudo calls from the LLMs and limited reading of certain env files.”Mark Arslan, Senior Software Engineer
- ImpactWhat does MintMCP let you do day to day?
- “If we see a trend toward a particular server, we bring security in, review it, and roll it out, because there's clearly value there. And when someone uses a random unofficial GitHub MCP, the dashboard lets us reach out: please switch to the official one, here's the one-line command, carry on.”Mark Arslan, Senior Software Engineer
- ImpactDid you consider building it yourselves?
- “There's always a build-versus-buy discussion. We could have self-hosted an open-source option, but maintaining it and adding all the features you get out of the box would be a lot more overhead. You can write it and ship it, but then you have to maintain it and make sure it works correctly, which is the other half of the battle.”Mark Arslan, Senior Software Engineer
- EnablementHow did AI spread beyond engineering?
- “I did a presentation on my use of AI, and the next question I got asked was, ‘Can you set me up with this?’ My first instinct was to sit down, install the tools, and show them all the command line stuff, and that made me pause, because that's not scalable. So we started a small experiment with a few people in the exec team using Claude Cowork. Initially they didn't find a lot of value. The key unlock was once we added MCPs and they got access to their data: there was a lot they were able to automate. Of all the AI usage in Coursera today, more than 25% comes from Cowork, from non-technical folks.”Mustafa Furniturewala, CTO
- ImpactWhat's your advice to other enterprise leaders?
- “Work with your security team, work with your InfoSec team to do it in the right way, so that the enablement happens at enterprise security scale and you're still compliant, doing all the right things to protect your data. The key message around enablement and leaning in is important because AI is changing a lot of workflows. Any function can benefit and securely use these sophisticated capabilities if they have the right kind of tools.”Mustafa Furniturewala, CTO
Move from LLMs running wild to governed by policy
See how MintMCP lets you enable AI across the company with the visibility and controls security needs. Read more stories