> ## Documentation Index
> Fetch the complete documentation index at: https://docs.splox.io/llms.txt
> Use this file to discover all available pages before exploring further.

# LLM endpoints

> Which providers an account can use, which models each one offers, and what those models accept

An agent names a model and the credential it runs on. These two endpoints are how
you find out which names are valid for your account before you write them into a
harness.

## The endpoints

```bash theme={null}
curl -s "$SPLOX_BASE_URL/v2/llm-endpoints" \
  -H "Authorization: Bearer $SPLOX_API_KEY"
```

```json theme={null}
{
  "data": [
    {"id": "019e2d78-0f8e-7f47-8c59-46bfa4f8f076", "name": "OpenAI",    "client": "openai",    "is_platform": true, "is_default": false},
    {"id": "019e2d78-0f8f-7fc5-b2a9-ed2b8ce9edbe", "name": "Anthropic", "client": "anthropic", "is_platform": true, "is_default": true},
    {"id": "019e2d78-0f8f-7581-9dbd-2adbfd795691", "name": "Gemini",    "client": "gemini",    "is_platform": true, "is_default": false},
    {"id": "019f0eb5-32a0-7291-bdc3-36441b60f86f", "name": "Splox",     "client": "splox",     "is_platform": true, "is_default": false}
  ]
}
```

The list is the platform's own endpoints plus any you have connected yourself.

<ResponseField name="id" type="uuid">
  What an agent writes as `text_llm_endpoint_id`. It is a raw UUID, not a public id: endpoints are referenced inside a harness's files, not addressed as API resources.
</ResponseField>

<ResponseField name="client" type="string">
  The provider slug — `anthropic`, `openai`, `gemini`, `splox` — which is also what an agent writes as `provider`.
</ResponseField>

<ResponseField name="is_platform" type="boolean">
  True when the platform holds the credential, so you need none of your own.
</ResponseField>

<ResponseField name="is_default" type="boolean">
  True for at most one endpoint: the one the server stamps into an agent that names none.
</ResponseField>

Naming an endpoint is optional. Leave `text_llm_endpoint_id` out of an agent and
it gets the default one — above, Anthropic. What you cannot do is leave the agent
with neither a provider nor an endpoint: nothing then says whose credential the
turn runs on, and the run fails on its first turn.

## The models of one endpoint

```bash theme={null}
curl -s "$SPLOX_BASE_URL/v2/llm-endpoints/019e2d78-0f8f-7fc5-b2a9-ed2b8ce9edbe/models" \
  -H "Authorization: Bearer $SPLOX_API_KEY"
```

```json theme={null}
{
  "data": [
    {
      "id": "claude-haiku-4-5-20251001",
      "name": "claude-haiku-4-5-20251001",
      "operation": "chat_completion",
      "capabilities": {
        "context_window": 200000,
        "modalities": ["text", "image"],
        "reasoning": {"control": "effort", "levels": ["off", "high"], "default_level": "high"},
        "supports_realtime": false,
        "supports_reasoning": true,
        "supports_streaming": true,
        "supports_tools": true
      },
      "input_schema": { "type": "object", "properties": { "...": "..." } }
    }
  ]
}
```

Models come back in display order, across every operation the provider serves —
`chat_completion`, `responses`, `realtime`. The Anthropic endpoint above answers
with eleven, all `chat_completion`, ids from `claude-haiku-4-5-20251001` to
`claude-opus-5`.

`id` is the value an agent writes as `model`:

```python theme={null}
assistant = agent(
    "Assistant",
    system_prompt="Answer in one sentence.",
    model="claude-sonnet-4-6",
    provider="anthropic",
)
```

Model choice is optional too — omit it and the server assigns its default.

## What a model accepts

`capabilities` is what the model can do; `input_schema` is the JSON Schema of the
knobs it accepts, which are the ones you may pass to the agent as
`additional_llm_config`. It is a real schema, with types, bounds and defaults:

```json theme={null}
{
  "temperature": {
    "type": ["number", "null"],
    "default": null,
    "maximum": 2,
    "minimum": 0,
    "description": "What sampling temperature to use, between 0 and 2"
  },
  "thinking_tokens": {
    "type": "integer",
    "default": 4096,
    "minimum": 1,
    "description": "Maximum number of tokens budgeted for hidden reasoning"
  }
}
```

The knobs differ per model, which is the point of asking: on this endpoint the
4.5-generation models take `thinking_tokens`, the 5-generation models take
`thinking_effort`, and only the Sonnet models offer `enable_1m_context`. Reading
the schema beats guessing at a parameter the provider will reject.

## Visibility

Platform endpoints are visible to everybody; an endpoint you connected is visible
only to you. An endpoint that belongs to somebody else and one that does not exist
answer the same way:

```json theme={null}
{"type":"https://api.splox.com/problems/not_found","title":"Not Found","status":404,"detail":"The resource does not exist or is not visible to the principal.","code":"not_found"}
```

## In the SDKs

<CodeGroup>
  ```python Python theme={null}
  endpoints = client.llm_endpoints.list()
  default = next(e for e in endpoints if e.is_default)
  models = client.llm_endpoints.models(default.id)
  print(default.client, [m.id for m in models][:3])
  # anthropic ['claude-fable-5-1', 'claude-opus-5', 'claude-fable-5']
  ```

  ```ts Node theme={null}
  const endpoints = await client.llmEndpoints.list();
  const defaultEp = endpoints.find((e) => e.isDefault)!;
  const models = await client.llmEndpoints.models(defaultEp.id);
  ```

  ```go Go theme={null}
  endpoints, err := client.LLMEndpoints.List(ctx)
  models, err := client.LLMEndpoints.Models(ctx, endpointID)
  ```
</CodeGroup>