kReative Labs
Back to kReative Labs

Live 2026

Nanda demo

A working prototype of the resolution flow from "NANDA: Networked Agent Name Discovery Architecture." A client that knows only an agent's name follows signed pointers to verified, actionable details, and a tampered field gets rejected by real cryptography, not a convention.

Open the guided demo

The problem

A name is not a trustworthy way to find an agent

As more software runs as autonomous agents, something has to answer a basic question safely: given only an agent's name, how does a client find where it lives, what it can do, what it costs, and whether any of that can be trusted, without taking the answer on faith? A phone-book lookup is not enough if the phone book itself can be altered.

The NANDA paper proposes a two-hop, signed-pointer architecture for exactly this. This project builds it far enough to be checked: real Ed25519 signatures, a genuine tamper test that fails closed, and an interactive walkthrough built so the mechanism is legible to someone seeing it for the first time.

Architecture

Two hops, then verify before you act

NANDA two-hop resolution A client knows only an agent name. Hop one queries the NANDA lean index and receives a signed AgentAddr pointer containing only a facts URL, key ID, TTL, and signature, deliberately with no capabilities, pricing, or trust data. Hop two follows that facts URL to the AgentFacts service, which returns signed metadata including capabilities, pricing, trust, and endpoints. The client verifies the signature with Ed25519 over canonical JSON before acting; changing any signed field, such as the trust score, makes verification fail. Client knows only an agent NAME NANDA index hop 1: name lookup returns signed AgentAddr facts_url, key_id, ttl, sig no capabilities / pricing / trust AgentFacts service hop 2: follow facts_url returns signed metadata capabilities, pricing, trust, endpoints signed, fetched separately Verify + act Ed25519 over canonical JSON one changed field fails
The AgentAddr stays deliberately thin. Heavy, fast-changing data such as pricing and trust lives in AgentFacts, signed and fetched on its own so it can change without touching the index.

Signed canonical JSON was chosen over W3C Verifiable Credentials for this level: VCs add JSON-LD context resolution and a proof suite that buy interoperability the prototype does not need yet, while canonical-JSON signing already gives full tamper detection. The upgrade path is to wrap the body in a VC envelope later.

Detail

At a glance

ResolutionTwo-hop: NANDA index → AgentAddr (signed pointer) → AgentFacts service → AgentFacts (signed metadata).
VerificationEd25519 signatures over canonical JSON via the platform Web Crypto API; any changed signed field fails verification.
Demo modesGuided one-click walkthrough, a driveable protocol trace with tamper / TTL-expiry / circuit-break / revoke scenarios, and a multi-agent DAG economy with per-agent billing.
Under the hoodTTL-aware LRU cache, binary max-heap adaptive router, Kahn's-algorithm DAG scheduler, circuit-breaker state machine, SHA-256 Merkle-chained audit log.
Also includedHeadless CLI demo and a 20-assertion test harness exercising the real registry and verification code, not a copy.
StackNext.js, deployed to Railway.
StageLive.