Everyone's building AI agents right now. Almost nobody's building the boring layer underneath them: how do agents find each other?
I hit this in my own stack. I run a few MCP servers one for search, one for my database, one exposing my product's API. Their addresses lived in a JSON config I edited by hand. When I added a fourth, the file got messy. When one went down, I found out because something downstream broke and I debugged backwards. There was no single source of truth for what agents existed or whether they were alive.
MCP gave agents a common language. There's still no discovery layer. That felt like having HTTP without a service registry.
So I built agentreg. It's a self-hosted service registry for agents one Go binary:
agentctl serve &
agentctl register search-agent -c search -e http://localhost:3000
agentctl find search # discover by capability, not a hardcoded URL
agentctl list # everything, with live health
It health-checks every registered agent on a heartbeat, so a dead one shows up as unhealthy
before it takes something down with it. That single feature has already saved me a debugging session.
A fair question I keep getting: "isn't this just service discovery, like Consul?" Close, but the query is different. Consul answers where is this service? a host and port. Agents need what can this thing do, is it healthy, and can I trust it? That's capability-first, and it's why the registry is agent-shaped rather
than micro service-shaped.
I'm being upfront about the stage: this is v0.1, I built it solo, and it does discovery + health and nothing more yet. No federation, no auth, no trust verification — though the code has a pluggable verifier seam for exactly that later. It's MIT, self-hosted, no cloud, no account.
Install:
brew install mkk2026/tap/agentreg
Repo: https://github.com/mkk2026/agentreg
If you run more than a couple of MCP servers, I'd genuinely like to know whether this solves a real annoyance for you or whether I'm a year early. The most useful thing you can tell me is whether you're still running it a week after you try it. That's the only feedback that counts.