Proposal: Implement A2A Agent Registrar for Curated Agent Discovery(Registry-Based)

[liuzengh]

I propose implementing a Agent Registrar to facilitate curated discovery of agents across organizations. This would provide a centralized mechanism for agent registration, discovery, and management while respecting organizational boundaries and visibility controls.

PR: #296

Benefits

• Enterprise-grade agent discovery with organizational boundaries
• Standardized registry interface for consistent agent discovery
• Controlled visibility for sensitive or internal-only agents
• Authentication model preserving security of agent registrations

[mukteshkrmishra]

+1 to this.

[edenreich]

Would be cool to have a centralized A2A agents registry that are highly optimized with security best practices in place (signature, container scanning etc) - that you can simply download and run. I vote for this. Maybe I'm thinking too much Kubernetes but would be nice to simply have a simple way (like ollama has) to download and run them. I'm working on a Kubernetes operator for the Inference Gateway if any one is interested in collab - let me know!

[ahoblitz]

Sharing Model Context Protocol had been working on a pre-release registry here: https://github.com/modelcontextprotocol/registry

IBM's Context Forge MCP Gateway has also been attempting to offer some integration with A2A so it may have some limited implementation of a rudimentary A2A gateway https://github.com/IBM/mcp-context-forge

[chorghemaruti64-creator]

The curated registry approach is valuable for trust, but there's a scale constraint worth thinking through: curation doesn't compose across organizations.

If Organization A's registry knows about their 50 agents, and Organization B has their own 50, there's no protocol for Agent X at Org A to discover Agent Y at Org B — unless both registries federate or both register in some shared namespace.

This is the distinction between:

• Directory (curated, per-org, human-maintained)
• DNS-style discovery (self-registering, federated, machine-readable resolution)

For the curated case, you need human review → quality guarantee but slow & fragmented.
For the DNS case, you need behavioral attestation → quality emerges from track record.

The implementation we've been building (Agenium) takes the second path — agents self-register with capability manifests, and trust accumulates via behavioral history at a canonical address. Think of it as complementary to curated registries: curated for high-trust production agents, open discovery for the long tail.

Happy to discuss the federation model if that's where this PR is heading.

[ThriveLabsAI]

FWIW, we've been operating a small public agent network on the opposite end of this spectrum — registry-free, P2P discovery — and it's worth holding both as complementary rather than competing.

The registrar model (this proposal) is right for curated, enterprise, boundary-respecting discovery. But it leaves a gap for the "first contact, no prior relationship, no one to ask" case: a curated registry still requires a human to approve the listing, and a directory you have to be added to.

The alternative primitive we use:

• Beacon — a short, agent-readable page listing your "current asks" (what you want other agents to do right now). No JS, no registry.
• Signed knock — a caller proves identity with its own Ed25519 key (identity = the key), no shared secret, no account to provision.

The effect is discovery without a directory: any agent that lands on the beacon can act immediately — no approval, no token. https://knock.thrivers.ai/beacon

Not arguing against a registrar — curated discovery is clearly needed for enterprise. Just flagging that the registry-free pattern covers the "cold, cross-org, first-contact" case that a curated list can't, and the two compose well (registrar for trusted orgs, beacon+knock for the open edge).

[edenreich]

Inference Gateway has implemented something similar to what you describe the catalog is hosted on your own GitHub repository, and you can pull A2A agents by just configuring them on the cli.

Or if you think others can benefits from them then you can also publish the agents in the global registry by referencing your agent's repo (public OSS agents).

https://registry.inference-gateway.com/agents/