Dex is a universal package manager for AI coding agent capabilities. You define skills, commands, rules, and MCP servers once in HCL, and Dex compiles them into whatever Claude Code, Cursor, and GitHub Copilot expect natively.
The problem
Every major AI coding tool has its own format for custom instructions, commands, and integrations. Teams end up duplicating the same intent across three different systems. When something changes, they update it in three places, or they forget and the configurations drift. The more agents you run across more projects, the worse it gets.
I built Dex because I was living this problem at Launch. We were rolling out AI coding tools across delivery teams and clients, and managing per-tool configurations by hand wasn't scaling.
Distribution was the other half of the problem. The existing marketplaces (Claude's, for example) assume technical users comfortable with GitHub. For the non-technical people on our teams, that was a barrier. And every seat needed a license, which added cost that didn't make sense when we just needed to share internal packages.
How it works
Dex treats agentic context and configuration as packages. A package defines skills, commands, rules, agents, and MCP servers in a single HCL file. Each resource type gets translated to the native format for the target platform (Claude commands, Copilot prompts, Cursor commands, etc.). If a resource type isn't supported on a platform, it gets skipped cleanly.
Packages can be hosted anywhere. Dex supports Git repositories, local filesystems, HTTPS registries, S3 buckets, and Azure Blob Storage. Teams pick whatever fits their infrastructure instead of being locked into a single marketplace.
Profiles let you target configurations to specific kinds of work. Instead of one massive setup that tries to cover everything, you scope capability sets to what a given task actually needs, which cuts down complexity and keeps the agent focused. Configuration lives in version control. New team members get the right setup from `dex sync` instead of following a wiki page that's three months out of date.