models-config.json tells the proxy-router how to translate on-chain modelId values into actual backend HTTP calls. The schema is enforced by proxy-router/internal/config/models-config-schema.json.

​
Per-model fields

FieldRequiredNotes
modelIdyesThe on-chain model id (32-byte hex)
modelNameyesHuman-readable name
apiTypeyesOne of openai, claudeai, prodia-sd, prodia-sdxl, prodia-v2, hyperbolic-sd, etc.
apiUrlyesFull backend URL including the endpoint path
apiKeyoptionalBackend API key (resale providers usually need this)
concurrentSlotsoptionalDistinct concurrent chats your backend can handle for this model
capacityPolicyoptionalidle_timeout or simple
concurrentSlots is per-model, not global. The proxy-router will accept up to sum(concurrentSlots) across all configured models concurrently. There is no built-in global cap. If your hardware can serve only one of N models at a time, register all N models but only post a bid for the one currently active, and rotate the bid when you want to switch. See Pausing or temporarily disabling a model offering.
| parameters | optional | Per-apiType extra parameters (e.g. SD cfg_scale, steps) | For the canonical per-apiType schema, see the JSON schema file linked above.

​
Examples

{
  "$schema": "./internal/config/models-config-schema.json",
  "models": [
    {
      "modelId": "0x0000000000000000000000000000000000000000000000000000000000000000",
      "modelName": "llama2",
      "apiType": "openai",
      "apiUrl": "http://localhost:8080/v1/chat/completions"
    },
    {
      "modelId": "0x0000000000000000000000000000000000000000000000000000000000000001",
      "modelName": "inference.sdxl.txt2img.v1",
      "apiType": "prodia-v2",
      "apiUrl": "https://inference.prodia.com/v2/job",
      "apiKey": "FILL_ME_IN"
    },
    {
      "modelId": "0x0000000000000000000000000000000000000000000000000000000000000002",
      "modelName": "SDXL1.0-base",
      "apiType": "hyperbolic-sd",
      "apiUrl": "https://api.hyperbolic.xyz/v1/image/generation",
      "apiKey": "FILL_ME_IN",
      "parameters": {
        "cfg_scale": "5",
        "steps": "30"
      }
    },
    {
      "modelId": "0x0000000000000000000000000000000000000000000000000000000000000003",
      "modelName": "claude-3-5-sonnet-20241022",
      "apiType": "claudeai",
      "apiUrl": "https://api.anthropic.com/v1/messages",
      "apiKey": "FILL_ME_IN"
    },
    {
      "modelId": "0x0000000000000000000000000000000000000000000000000000000000000004",
      "modelName": "gpt-4o-mini",
      "apiType": "openai",
      "apiUrl": "https://api.openai.com/v1/chat/completions",
      "apiKey": "FILL_ME_IN"
    },
    {
      "modelId": "0x0000000000000000000000000000000000000000000000000000000000000005",
      "modelName": "text-embedding-bge-m3",
      "apiType": "openai",
      "apiUrl": "https://api.venice.ai/api/v1/embeddings",
      "apiKey": "FILL_ME_IN"
    },
    {
      "modelId": "0x0000000000000000000000000000000000000000000000000000000000000006",
      "modelName": "tts-kokoro",
      "apiType": "openai",
      "apiUrl": "https://api.venice.ai/api/v1/audio/speech",
      "apiKey": "FILL_ME_IN"
    },
    {
      "modelId": "0x0000000000000000000000000000000000000000000000000000000000000007",
      "modelName": "whisper-1",
      "apiType": "openai",
      "apiUrl": "https://api.openai.com/v1/audio/transcriptions",
      "apiKey": "FILL_ME_IN"
    }
  ]
}

​
Inline content (for TEE / Akash)

For deployments where mounting a JSON file is impractical, set MODELS_CONFIG_CONTENT to a single-line JSON string with the exact same shape as above. The proxy-router reads it instead of MODELS_CONFIG_PATH. This is the standard approach in TEE and Akash deployments.