LaunchDarkly review: features, pricing, and AI capabilities

Platform overview

LaunchDarkly's AI surface is AgentControl. Configs pull prompts, instructions and model settings out of application code, in completion or agent mode, with variations, targeting, monitoring, experimentation and judges attached for online evaluation.

Agent Optimization entered public beta in July 2026. It runs optimization passes across models, prompts and configurations, scores them with a judge model, and deploys the winner from LaunchDarkly. AgentControl monitoring can now be filtered and grouped by context kind, so AI cost, latency, tokens and satisfaction can be attributed to individual tenants or users.

Agents drive it through a hosted OAuth MCP server at https://mcp.launchdarkly.com/mcp/launchdarkly, covering feature management, AgentControl and observability, with a local npx server for federal and EU instances. LaunchDarkly also publishes installable agent skills for multi-step workflows such as flag cleanup through to a pull request.