Integrations

A tool that speaks OpenAI's chat completions API works with AI Tokens. It needs three values:

Tools give the base URL different names: Base URL, API Base, api_url, baseURL, OPENAI_API_BASE. The value is the same everywhere, and it ends in /v1, without /chat/completions.

Where a tool asks for a provider type, choose OpenAI Compatible. Where it offers only OpenAI, choose it and change the base URL.

Tested, or from the documentation#

Every page says which of the two it is.

• Tested: we ran the example on the live API on the date given, with the version of the tool named on the page, and the page shows what came back.
• From the documentation: the settings come from the tool's own documentation, read on the date given. We did not run the tool. If a setting does not work as written, tell us at info@daikolab.ch.

Which model#

The catalogue lists the models AI Tokens serves today, with their prices and the country where each one runs. This table adds what a tool needs to be told, as we served it on 11 October 2026.

Tools is what we checked on 11 October 2026: one request with a tool on each model, with the OpenAI SDK, and the same with stream: true on the first four rows. Each model returned the call in OpenAI's shape. A coding agent needs a model with tools.

Context window. Set it in the tool to the number in this table. A tool that does not know it assumes its own default, and may send a conversation longer than the model can read.

Maximum output. Set it to 8,192 tokens or less. The API refuses a larger max_tokens with 422.

Reasoning#

qwen3.8-27b and gpt-oss-120b reason before they answer. The reasoning is not returned: tools that show a model's thinking show nothing for these two. It is counted in usage.completion_tokens, billed as output, and it counts against max_tokens. With a low cap the answer comes back cut or empty, with finish_reason: "length".

On qwen3.8-27b, "reasoning_effort": "none" switches reasoning off for a request. Asked for one sentence about the Gotthard tunnel on 11 October 2026, it used 176 output tokens with reasoning and 47 without. Each tool page shows how to send the field, where the tool allows it. More in Every parameter.

auto#

With "model": "auto", the API chooses the model for each request: today qwen3.6-35b for routine work, and qwen3.8-27b for work that needs reasoning or a long input. It suits chat interfaces and workflows, such as Open WebUI and n8n. Where a tool asks for the context window of auto, give it 32,768, the smaller of the two.

For coding agents, name a model. auto keeps a task on the model it started with, so that the agent does not change model halfway through. A task that started on qwen3.6-35b stays within its 32,768 tokens, even when the files the agent reads grow past them.

What the API does not do#

• Anthropic's Messages API. /v1/messages answers 404 with unknown_endpoint. Tools that speak only Anthropic's format, such as Claude Code, cannot connect directly.
• Fill-in-the-middle. /v1/completions passes the prompt to the model as a chat message, and refuses suffix. Tab completion built on fill-in-the-middle does not work here: Continue's autocomplete, Zed's edit predictions. Chat, edits and agents do.
• Tools and streaming on the Responses API. /v1/responses answers 501 to stream: true. On 11 October 2026 it also did not read a tool result sent back as function_call_output: the model asked for the same tool again. Tools that can use either API have to use chat completions; the pages of n8n and the OpenAI Agents SDK show the setting.
• Parameters it does not know. A field that is not in Every parameter gets a 400 that names it. Some tools send fields of other engines when you fill in their settings, such as top_k, min_p, repetition_penalty or prompt_cache_key. Leave those settings empty.
• Embeddings from other models. /v1/embeddings always answers with bge-m3: 1,024 dimensions, at most 32 texts per request. See Embeddings.

/v1/models lists every model AI Tokens serves, and some of them do not chat: bge-m3, bge-reranker-v2-m3, whisper-1, whisper-1-hd, tts-1, and the picture and video models where there are any. When a tool fills its model menu from that list, choose a chat model.

Headers#

Where a tool lets you add HTTP headers, X-Siati-Tier sets the tier of its requests. The zone where a request ran comes back in the X-Zona header of the answer.