Peer Discovery

The discovery protocol uses a DHT network (built on BEP 5) as a decentralized directory of seller nodes, combined with an HTTP metadata endpoint for retrieving provider details and Skills. All nodes bootstrap through dedicated AntSeed infrastructure.

DHT Topic Hashing

Sellers announce multiple topic types. Each topic is SHA1-hashed for DHT lookup.

Topic Type	Plain Topic String	Key Normalization
Provider	antseed:{provider}	trim + lowercase
Model (canonical)	antseed:service:{model}	trim + lowercase
Model (search fallback)	antseed:service-search:{model}	trim + lowercase, then remove spaces, -, _ (keep .)
Capability	antseed:{capability} or antseed:{capability}:{name}	trim + lowercase

model-search topics are announced only when the compact key differs from the canonical key.

Example:

• kimi 2.5 -> canonical antseed:service:kimi 2.5, search antseed:service-search:kimi2.5
• kimi-2.5 -> canonical antseed:service:kimi-2.5, search antseed:service-search:kimi2.5
• kimi_2.5 -> canonical antseed:service:kimi_2.5, search antseed:service-search:kimi2.5

Buyer model discovery queries canonical model topic first, then also queries model-search when keys differ.

Bootstrap Nodes

Host	Port
dht1.antseed.com	6881
dht2.antseed.com	6881

DHT Configuration

Parameter	Value
Port	6881
Re-announce interval	15 minutes
Operation timeout	10 seconds

Metadata Endpoint

Each seller runs an HTTP server exposing GET /metadata which returns JSON-serialized PeerMetadata with pricing, capacity, and optional metadata tags/protocol hints.
By default, metadata is fetched from http://{host}:{port}/metadata (metadataPortOffset = 0).

PeerMetadata

metadata structure

{
  "peerId": "a1b2c3d4...40 hex chars (EVM address)",
  "version": 5,
  "displayName": "Acme Inference - us-east-1",
  "publicAddress": "peer.example.com:6882",
  "providers": [{
    "provider": "anthropic",
    "services": ["claude-sonnet-4-6", "claude-haiku-4-5"],
    "defaultPricing": {
      "inputUsdPerMillion": 3,
      "cachedInputUsdPerMillion": 0.3,
      "outputUsdPerMillion": 15
    },
    "servicePricing": {
      "claude-sonnet-4-6": { "inputUsdPerMillion": 3, "cachedInputUsdPerMillion": 0.3, "outputUsdPerMillion": 15 },
      "claude-haiku-4-5": { "inputUsdPerMillion": 1, "cachedInputUsdPerMillion": 0.1, "outputUsdPerMillion": 5 }
    },
    "serviceCategories": {
      "claude-sonnet-4-6": ["coding", "privacy"]
    },
    "serviceApiProtocols": {
      "claude-sonnet-4-6": ["anthropic-messages"]
    },
    "maxConcurrency": 5,
    "currentLoad": 2
  }],
  "region": "us-east",
  "timestamp": 1708272000000,
  "signature": "eip191...130 hex chars"
}

Recommended category tags: privacy, legal, uncensored, coding, finance, tee (custom tags are allowed).

publicAddress is optional. When present, buyers should prefer it over the raw host learned from the DHT announcement. This is intended for deployments where DHT traffic exits from one IP but buyers must connect to another address, such as a Kubernetes load balancer.

Peer Scoring

Dimension	Weight	Description
Price	0.30	Lower price scores higher (inverted min-max)
Latency	0.25	Lower latency scores higher (EMA-based)
Capacity	0.20	More available capacity scores higher
Reputation	0.10	Higher reputation scores higher (0-100)
Freshness	0.10	Recently seen peers score higher
Reliability	0.05	Lower failure rate and streak scores higher

All factors are min-max normalized across the eligible candidate pool. Peers below minPeerReputation (default: 50) are excluded before scoring. Peers in a failure cooldown (exponential backoff) are also excluded.

Buyers can filter by capability, Skill, minimum reputation, and price ceiling.