Public Gateway

The public gateway provides rate-limited, anonymous access to the mycellm network. No API key required.

POST /v1/public/chat/completions

Same request format as /v1/chat/completions with restrictions:

• Rate limit: 5,000 tokens/day per IP, 10 requests/minute
• Max tokens: 1,024 per response
• Max message length: 2,000 characters
• Model: auto-selected (best available), user cannot choose
• Metadata: stripped — no peer IDs, routing details, or credit info

Request

{
  "messages": [{"role": "user", "content": "Hello"}],
  "stream": true
}

Streaming metadata

The gateway includes mycellm metadata in streaming chunks:

// First chunk — node attribution
{"mycellm": {"node": "99e58f4c", "served_by": "mycellm-public"}}

// Final chunk — latency
{"mycellm": {"node": "99e58f4c", "latency_ms": 485, "served_by": "mycellm-public"}}

Error responses

Status	Meaning
429	Rate limit exceeded
503	No models available
400	Invalid request

GET /v1/public/models

What the public gateway can actually serve right now, with the tier of each. No authentication.

{
  "object": "list",
  "data": [
    { "id": "Qwen3.6-35B-A3B", "tier": "capable", "vision": false },
    { "id": "Qwen3.5-9B",      "tier": "fast",    "vision": false }
  ],
  "tiers": { "frontier": 0, "capable": 1, "fast": 2, "tiny": 2 }
}

tiers counts how many models satisfy each floor — a floor admits its own tier and everything above it, which is why a 35B counts toward both capable and fast. A model whose size cannot be determined has no tier and counts toward none: unknown means unknown, never “probably fine”.

Why this is not /v1/models

/v1/models answers what does this node know about. This answers what will answer me if I press send — derived from the gateway’s own candidate list, so an unapproved fleet node, a peer that dropped its connection, or a vision-only backend is excluded. Those two lists diverge constantly, and a model picker populated from the wrong one offers choices that can only fail.

Asking for a specific model or tier

POST /v1/public/chat/completions accepts either — never both:

# a quality floor: the gateway chooses, but not below this tier
curl -X POST https://api.mycellm.dev/v1/public/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"messages":[{"role":"user","content":"hi"}],
       "mycellm": {"min_tier": "capable"}}'

# or one named model
curl -X POST https://api.mycellm.dev/v1/public/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen3.5-9B","messages":[{"role":"user","content":"hi"}]}'

A floor that nothing meets is refused, not quietly downgraded:

{ "error": {
    "message": "No frontier-tier model is available right now (2 model(s) reachable below that tier).",
    "type": "no_candidate",
    "requested_tier": "frontier" } }

The no_candidate type matters to clients: it is a 503 that retrying will not fix, unlike the busy-503 above it. Backing off four times before reporting “the network is down” is wrong when the network is up and simply does not hold that model.

GET /v1/node/public/stats

Network stats — no auth required.

{
  "network_name": "mycellm-public",
  "nodes": {"total": 3, "online": 3, "seeding": 3},
  "compute": {"total_tps": 12.5, "total_vram_gb": 80.0, "total_ram_gb": 87.6},
  "models": {
    "unique": 2,
    "names": ["Qwen2.5-3B-Instruct-Q8_0", "Mistral-Small-24B"],
    "by_tier": {"tier1": [...], "tier2": [...], "tier3": []}
  },
  "activity": {"total_requests": 1542, "total_tokens": 89420},
  "top_contributors": [{"name": "aurora", "tps": 12.3, "models": 2}]
}