From Registry to Index: Our Journey to Production NANDA

Where We Started

When we launched KnowYourModel, the goal was straightforward: give AI agents a verifiable identity. Register an agent, certify its capabilities through automated trials, and let consumers look it up by name. A yellow pages for AI agents.

We built this on Cloudflare Workers — a global edge runtime that gives us sub-millisecond cold starts and D1 for structured data. Registration, search, and lookup worked. Certification trials graded agents against real tasks. Compliance policies evaluated agent behavior. The trust infrastructure was solid.

But as the agent ecosystem grew, we realized a registry wasn't enough. Agents don't just need to be listed. They need to be discovered — dynamically, contextually, across organizational boundaries. That's when the NANDA Index paper from MIT Media Lab landed, and it described exactly what we needed to build.

The NANDA Vision

NANDA — the Naming and Discovery Architecture for AI Agents — reimagines agent discovery as a three-layer system, analogous to DNS but purpose-built for the trust, capability, and protocol richness that AI agents require.

Layer 1: Lean Index

Lightweight AgentAddr records (≤120 bytes) — cryptographically signed pointers with Ed25519 signatures. Think DNS A records, but with built-in verification and metadata pointers.

Layer 2: Rich Metadata

AgentFacts documents — full capability declarations, trust certifications, performance metrics, content flags. Hosted by the agent, verified by the index.

Layer 3: Adaptive Resolution

Context-aware endpoint selection — scoring candidates on geography, trust, capability match, and health data. Not just "where is it?" but "which one is best right now?"

What We Built

Implementing the NANDA Index on our existing Cloudflare Workers infrastructure turned out to be a natural fit. Our registry tables became the data layer for AgentAddr records. Our KV cache became the resolution cache. Our existing certification and compliance systems became the trust data sources for AgentFacts v2.

The key additions:

AgentAddr records — Lean, Ed25519-signed pointers to each agent's metadata. Every registration now generates a signed AgentAddr, cached at the edge in Cloudflare KV.
AgentFacts v2 — Upgraded from plain JSON to W3C Verifiable Credential envelopes. Capabilities, trust certifications, performance data, and content flags — all cryptographically verifiable.
Adaptive Resolver — Three resolution strategies (static, rotating, adaptive) with a composite scoring engine that evaluates geography, trust, capability match, and health data.
Protocol Switchboard — Bidirectional adapters for A2A, MCP, and AGNTCY/OASF. Register once in any format, discoverable across all protocols.
CRDT gossip — LWW-Register conflict-free replication with gossip protocol for federation with other NANDA Index nodes in the global mesh.

The Bridge Matters Most

If we had to pick the single most impactful addition, it's the Protocol Switchboard. The agent ecosystem in 2026 is genuinely multi-protocol — Google's A2A for agent communication, Anthropic's MCP for tool access, Cisco's AGNTCY for enterprise orchestration. Each has its own discovery silo.

Our Switchboard breaks down these walls. An A2A agent registered in Google's ecosystem becomes discoverable by MCP clients querying KnowYourModel. An MCP tool server gains AgentFacts-level trust metadata. An AGNTCY agent joins the federated mesh. The adapters handle the translation transparently — the agent doesn't need to change anything.

This is what makes the NANDA vision practical: not a new protocol to replace A2A or MCP, but a discovery layer that sits above all of them and makes every agent findable by everyone.

What's Next

We're now a functioning node in the NANDA Index mesh, but the work is just beginning. The areas we're focused on next:

Agentic SafeSearch

Content-policy filtering at the resolution layer. Query only agents that match your content requirements — kid-safe, HIPAA, jurisdiction-specific.

ZTAA Integration

Zero-Trust Agent Architecture — continuous trust evaluation where every agent interaction is verified, not just the initial handshake.

Mesh Expansion

More federation peers. As other organizations deploy NANDA Index nodes, the discovery mesh grows — and every node benefits from the expanded agent catalog.

Enterprise Deployment

Private NANDA Index nodes for enterprise environments with controlled federation and on-premise agent discovery.

Learn More

For the deep technical details, check out the companion posts on the Nexartis blog:

Building DNS for AI Agents — AgentAddr records, Ed25519 signing, and the Lean Index
The Protocol Bridge — Switchboard adapters for A2A, MCP, and AGNTCY
Adaptive Resolution — Multi-strategy scoring and smart endpoint routing
CRDT Gossip — Federation sync with conflict-free replicated data types

And the foundational research: