LLM Gateways

A gateway sits between your app and api.ablatic.ai and gives you routing, failover and usage tracking across providers. Ablatic speaks the OpenAI API, so any gateway that supports a custom OpenAI-compatible provider works.

Bifrost

Two settings decide whether the integration works:

Base URL without /v1. Set base_url to https://api.ablatic.ai. Bifrost appends /v1/chat/completions itself. If you paste the URL you use with curl or the OpenAI SDK (https://api.ablatic.ai/v1), Bifrost calls /v1/v1/chat/completions and gets a 404. This is the most common failure: curl works, the gateway does not.
Attach the model to the key. Bifrost only routes to keys whose models list contains the requested model. Leave the list empty and Bifrost answers no keys found that support model: talos-preview without ever calling Ablatic.

Configuration

{
  "providers": {
    "ablatic": {
      "keys": [
        {
          "value": "sk-ablatic-...",
          "models": ["talos-preview", "talos"],
          "weight": 1.0
        }
      ],
      "custom_provider_config": { "base_provider_type": "openai" },
      "network_config": { "base_url": "https://api.ablatic.ai" }
    }
  }
}

The same two rules apply in the Bifrost web UI: enter the base URL without /v1, and assign the model to the key.

Requests

Prefix the model with your provider name. Bifrost strips the prefix and forwards talos-preview to Ablatic.

curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ablatic/talos-preview",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

Streaming works unchanged: set "stream": true and read the Server-Sent Events, the same as calling Ablatic directly.

Troubleshooting

Symptom	Cause	Fix
`404 Not Found` from the provider	`/v1` in the base URL, so the gateway calls `/v1/v1/...`	Set `base_url` to `https://api.ablatic.ai`
`no keys found that support model: ...`	The key's `models` list is empty	Add `talos-preview` (and any alias you use) to the key
`401` with code `invalid_api_key`	The key is wrong or truncated	Check the key in the console; it starts with `sk-ablatic-`

LiteLLM

LiteLLM uses the opposite base-URL convention from Bifrost: api_base includes the /v1.

import litellm

resp = litellm.completion(
    model="openai/talos-preview",
    api_base="https://api.ablatic.ai/v1",
    api_key="sk-ablatic-...",
    messages=[{"role": "user", "content": "Hello"}],
)

The openai/ prefix tells LiteLLM to speak the OpenAI protocol; it forwards talos-preview to Ablatic. The same values work in a LiteLLM proxy config.yaml under litellm_params. Cost tracking works out of the box: every response and the final stream chunk carry usage.

Other gateways

The same pattern applies to most gateways with a custom OpenAI-compatible provider: check whether the gateway appends /v1/... to the base URL itself (Bifrost) or expects it in the base URL (LiteLLM), and make sure the model is registered with the provider. Errors on /v1/chat/completions use the standard OpenAI error object, so gateways surface Ablatic error messages directly (see Errors).