AI APIs
The platform exposes its AI capabilities as HTTP APIs you can call straight from your own software. Two surfaces cover most integrations:
- Model Gateway — one OpenAI-compatible endpoint for every model on the platform: chat, embeddings, reranking, and the Responses API. Whether a model runs in your own cluster or at an external provider, you call the same base URL and the gateway routes the request.
- RAG — a deployable retrieval system that grounds answers in your own documents. Each RAG has its own API on its own deployment-specific URL, used to ingest and manage documents and retrieve relevant context.
Authentication
Every request carries an API key as a bearer token:
Authorization: Bearer cm_api_…
The key's grants decide which models and RAGs it can reach. Each service's portal view shows its base URL alongside a Create API key link.
Calling a model
The Model Gateway speaks the OpenAI API, so existing clients and SDKs work unchanged — point them at the gateway and pass your key.
from openai import OpenAI
client = OpenAI(base_url="https://api.<your-domain>/v1", api_key="cm_api_…")
resp = client.chat.completions.create(
model="qwen3-6-27b-fp8",
messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)
GET /v1/models lists the models your key can use. See Model Gateway for routing, model capabilities, and filtering.
Agentic workflows and tools
The gateway's Responses API (/v1/responses) runs a built-in tool loop — hand the model a task and a set of tools, and it drives the calls until it has a final answer, with no orchestration to host. Tools are provided over MCP, shared with the built-in chat: RAG is available out of the box, and you can add your own MCP servers.