agentreg

DNS for AI agents.
A self-hosted registry that lets your agents find each other by capability — with health and trust built in.

Everyone is building AI agents. Nobody is building the layer that lets agents exist alongside each other — discovery, health, trust. The moment you run more than a couple of MCP servers, you're hardcoding endpoints in config files, finding out an agent is down only when something breaks, and trusting every endpoint blindly.

MCP gave agents a common language. agentreg gives them a place to find each other.

Consul tells you where a service is. agentreg tells you what an agent can do, whether it's healthy, and (soon) how much to trust it.

One Go binary. Self-hosted. No cloud, no account, no cluster. Runs on your laptop at 3 agents and scales with you.

Install

# Homebrew (macOS / Linux)
brew install mkk2026/tap/agentreg

# or download a binary from the latest release
#   https://github.com/mkk2026/agentreg/releases

# or from source (Go 1.26+)
go install github.com/mkk2026/agentreg@latest

Quickstart

# 1. run the registry
agentctl serve &

# 2. register your agents
agentctl register search-agent -c search     -e http://localhost:3000
agentctl register db-agent     -c db-read     -e http://localhost:3001
agentctl register product-api  -c api,email   -e http://localhost:3002

# 3. see everything, with live health
agentctl list

# 4. discover by capability — use this in your agent's startup instead of a hardcoded URL
agentctl find search
agentctl find search --format json

agentctl list:

NAME          CAPABILITIES  STATUS     SOURCE  LAST HEARTBEAT  ENDPOINT
db-agent      db-read       unhealthy  local   just now        http://localhost:3001
product-api   api,email     healthy    local   just now        http://localhost:3002
search-agent  search        healthy    local   just now        http://localhost:3000

The daemon health-checks every registered agent on an interval, so a down agent shows up as unhealthy before it breaks something downstream. That's the difference between reading a status table and debugging a production incident backwards.

Why agentreg

• Capability-first discovery. Ask for what you need (find search), not where it lives. Endpoints move; capabilities don't.
• Health built in. Every agent is probed on a heartbeat. No separate monitoring stack to wire up.
• Self-hosted, zero cloud. One static binary. Your registry, your data, your network. Nothing leaves your machine.
• Delightful at 3 agents, load-bearing at 30. Cheap enough to adopt before you need it, so it's already there when you do.
• Built to grow into trust. The verification layer is pluggable from day one (see Roadmap).

How it works

  agentctl (CLI)  ──HTTP──▶  agentreg daemon
                                  │
                                  ├── Store     in-memory + atomic JSON persistence
                                  └── Verifier  health probe, on a heartbeat loop

Two design seams make the roadmap additive instead of a rewrite:

• Source field on every record (local today; peer:<id>, from:ans, from:mcp_registry later) so federation and cross-registry ingestion don't require a migration.

• Verifier interface — v1 ships a health probe; ANS-identity, prompt-injection, and tools/list verifiers plug into the same seam.

• labels — optional operator metadata (env, team, version, …) that rides along on every record and shows up in list. Set it via the API today:

  hljs language-bash

  Copy

  curl -X POST http://localhost:8080/agents -H 'Content-Type: application/json' \
    -d '{"name":"corebrim-search","capabilities":["search"],
         "endpoint":"http://10.0.0.12:3000",
         "labels":{"env":"prod","team":"platform","version":"1.4.2"}}'
  

agentreg vs. the alternatives

agentreg isn't trying to out-spec the standards bodies — it's the tool developers actually run, built to interoperate with them.

HTTP API

Configuration

agentctl serve
  --port 8080                        # listen port
  --store ~/.agentreg/registry.json  # persistence file
  --heartbeat-interval 15s           # how often to health-check agents
  --probe-timeout 3s                 # per-agent probe timeout

Roadmap

agentreg ships as the smallest lovable loop first, then grows toward being the trust fabric agents rely on.

• v0.1 — the loop (now). register · list · find · heartbeat, single binary.
• Next — trust. Pluggable verifiers: ANS identity, tools/list validation, prompt-injection checks. Verification-as-a-service.
• Next — federation. agentctl federate across registries; the Source field lights up.
• Later — the fabric. One client that discovers and verifies across every registry (ANS, MCP Registry, local), so agents transact across org boundaries safely.

Delightful extras on the list: a live list TUI, agentctl doctor, agentctl graph, and zero-config auto-register from an MCP manifest.

Contributing

Early days, and help is welcome. Open an issue to discuss a change, or send a PR.

git clone https://github.com/mkk2026/agentreg
cd agentreg
go test ./...
go build -o agentctl .

License

MIT © Core Brim Tech

_{Built by Core Brim Tech · agents deserve a substrate.}