How LaunchDarkly extended runtime control to the AI era with MCP
LaunchDarkly
Dev tools
/Speakeasy
As the leading feature management and runtime control platform, LaunchDarkly has always been about giving developers control over their software releases. The rise of AI agents has fundamentally changed the way people build software and increased the need for runtime control – and feature flags remain one of the most important tools for achieving it. The question became: how do you give agents the same level of sophisticated control over feature flags that human developers have?
That’s the challenge LaunchDarkly faced. The answer would require building something entirely new – an MCP server that could bridge the gap between AI agents and the powerful feature management capabilities of LaunchDarkly.
Robots need runtime control too
LaunchDarkly has spent years perfecting the art of feature management. Their platform allows developers to deploy code safer behind feature flags, gradually roll out features to specific user segments, and instantly kill switches when things go wrong. It’s a system built for human decision-making, with dashboards, alerts, and approval workflows designed around how engineering teams actually work.
AI agents are already shipping features end to end: writing code, running tests, and deploying changes. But the feature management layer hasn’t kept pace. Agents had no native way to create flags, configure targeting rules, or manage rollouts through the same workflows that human developers rely on.
“The potential was obvious, imagine an AI agent that could not only build a new feature but also create the appropriate feature flag, set up the targeting rules, and even monitor the rollout.”
Benjamin Woskow,
Senior Director of Engineering
They decided to build an external MCP server so that customers’ AI agents could participate in feature management workflows. One use case in particular made the opportunity concrete: the newest product at LaunchDarkly, AgentControl.
AgentControl applies runtime control principles specifically to AI agents in production, using feature-flagging techniques to control agent behavior, route across models for different user segments, and instantly mitigate or roll back unsafe behavior based on judge evaluations or other observed metrics. The MCP server is a natural delivery mechanism. Customers can integrate AgentControl into their MCP-powered workflows so agents can participate in governed rollout, evaluation, and intervention loops without leaving the systems they already use. Feature flagging expertise extended to the very tools that AI agents rely on.
Why Speakeasy
Building an MCP server isn’t the hard part; almost any team can scaffold one with AI-assisted tooling. The hard part is getting it right and keeping it right. The MCP protocol is rapidly evolving, best practices are still being established, and the difference between a working MCP server and a production-grade one is significant.
LaunchDarkly wasn’t looking for a partner to build servers. They were looking for a partner with deep expertise in the protocol. A team that stays on top of spec changes, incorporates best practices as they emerge, and handles the ongoing maintenance so LaunchDarkly doesn’t have to.
“We realized that building and maintaining a production-grade MCP server could detract from our ability to focus on delivering our core business value. We wanted to partner with someone who could accelerate our time to market and minimize our overhead costs.”
Benjamin Woskow,
Senior Director of Engineering
Speakeasy fit that model. LaunchDarkly could point Speakeasy at their existing OpenAPI specification and get a production-ready MCP server in return. When LaunchDarkly’s APIs change, the MCP server updates automatically. When the MCP protocol itself evolves, Speakeasy incorporates the latest spec changes and best practices without LaunchDarkly lifting a finger.
But the API-based generation is just the starting point. With the foundation in place, the LD team is iterating beyond their API: adding custom tools, prompts, and workflows that tie together multiple API endpoints.
The internal MCP story
The first user of the LaunchDarkly MCP server was the LaunchDarkly developer team itself. As the team dogfooded the MCP server internally, they realized how powerful MCP servers were for their own engineering workflows.
The flag cleanup use case hit especially close to home. LaunchDarkly’s own codebase had accumulated its share of stale flags, and the MCP server gave their engineers a way to identify and remove them using AI agents. The team built custom prompts that check flag status, analyze usage statistics, examine code references, and determine if a flag is safe to remove. Engineers started using this workflow daily.
This internal experience changed how LaunchDarkly thought about the Speakeasy partnership. What started as a way to ship an external MCP server quickly became a broader platform play. The team saw firsthand how MCP could transform internal developer productivity, and that led them to pursue the full Speakeasy MCP platform. not just for the customer-facing MCP server, but as a foundation for how their engineering team connects AI to the rest of its internal data sources.
Looking toward an AI-integrated future
The LaunchDarkly MCP server represents more than just a new integration; it’s a glimpse into how feature management will work in an AI-powered development world. What started as an external product for customers became a catalyst for internal transformation, proving that the same MCP infrastructure can serve both audiences.
The partnership with Speakeasy has allowed LaunchDarkly to be early pioneers in this space. As AI agents become more sophisticated and take on larger roles in software development, they’ll need the same level of access to operational controls that human developers have. LaunchDarkly, built with Speakeasy, is ready for that future.
