Register a model

Instead of embedding model configuration in every agent, register a model once and reference it by model_ref. This makes credential rotation and model upgrades easier across many agents.

Register a model with the CLI

Create a model registration JSON file.

{
  "model_id": "gpt-4o-mini",
  "provider_id": "openai",
  "params": {
    "model": "gpt-4o-mini",
    "temperature": 0.0,
    "api_key": { "$secretRef": "RESPONSE_PROVIDER_API_KEY" }
  }
}

Add it to an agentspace.

alquimia registry models add gpt-4o-mini.json --namespace default

Verify the registration.

alquimia registry models list --namespace default

Register a model programmatically

from alquimia.core.registry import Registry
from alquimia.core.models import ModelRegistration

registry = Registry(registry_dir="/var/lib/alquimia/registry")

registry.add_model(
    "default",
    ModelRegistration(
        model_id="gpt-4o-mini",
        provider_id="openai",
        params={
            "model": "gpt-4o-mini",
            "temperature": 0.0,
            "api_key": {"$secretRef": "RESPONSE_PROVIDER_API_KEY"},
        },
    ),
)

Use a registered model in an agent spec

{
  "assistant_id": "support-bot",
  "response": {
    "provider_id": "alquimia",
    "config": {
      "model_ref": "gpt-4o-mini"
    },
    "profile": {
      "system_prompt": "You are a helpful support assistant.",
      "evaluation_strategy": { "evaluation_strategy_id": "one-shoot" }
    }
  }
}

Per-use params override the registered defaults:

{
  "response": {
    "provider_id": "alquimia",
    "config": {
      "model_ref": "gpt-4o-mini",
      "params": {
        "temperature": 0.7,
        "max_tokens": 2048
      }
    },
    "profile": { ... }
  }
}

Use a registered model in a shield

{
  "shields": {
    "intent-classifier": {
      "provider_id": "shield-config",
      "connector": {
        "provider_id": "alquimia",
        "profile": {
          "system_prompt": "Classify user intent.",
          "evaluation_strategy": {
            "evaluation_strategy_id": "one-shoot",
            "structured_output": {
              "method": "json_schema",
              "json_schema": {
                "type": "object",
                "properties": { "intent": { "type": "string" } },
                "required": ["intent"]
              }
            }
          }
        },
        "config": { "model_ref": "gpt-4o-mini" }
      },
      "action": "observe"
    }
  }
}

Supported provider IDs

As of alquimia-core 0.5.3, params are validated against a per-provider schema at registration time and again when the merged configuration is instantiated. Passing a parameter that belongs to a different provider — for example, groq_api_key to an openai-compatible connector — is rejected with an error that names the offending keys and the provider they belong to.

Provider	Typical params
`openai`	`model`, `api_key`, `base_url`, `temperature`, `max_tokens`, `timeout`, `top_p`, `max_retries`, `organization`, `reasoning`, `reasoning_effort`, `output_version`, `store`, `extra_body`
`openai-compatible`	Same as `openai`; use `reasoning` for the Responses API. Add `harmony_preamble_as_reasoning: true` for gpt-oss/vLLM streaming to keep chain-of-thought out of the answer (0.5.5+)
`groq`	`model`, `api_key`, `groq_api_key`, `base_url`, `temperature`, `max_tokens`, `timeout`, `top_p`, `max_retries`, `reasoning_format`, `reasoning_effort`, `output_version`
`huggingface`	`model`, `api_key`, `base_url`, `temperature`, `max_tokens`, `timeout`, `top_p`, `max_retries`
`ollama`	`model`, `base_url`, `temperature`, `max_tokens`

$secretRef markers inside params are accepted at registration and resolved at inference time, including LOCAL-scoped secrets when the registration is loaded for a specific assistant.

As of alquimia-core 0.5.2, you can also register audio models (AudioModelRef, STTModelRef, TTSModelRef) and reference them from the agent spec’s audio field as audio_model_ref, stt_model_ref, and tts_model_ref.

See the Agent specification reference for the model_ref schema and Audio adapters for audio model refs.