Providers

10xGraph talks to model providers through a unified Agent interface. It has built-in clients for three provider families: OpenAI, Google (Gemini API and Vertex AI) and Anthropic (Claude Messages API, Vertex AI and Amazon Bedrock). Anything that speaks the OpenAI API, such as Ollama, vLLM or OpenRouter, works through the OpenAI client with a base_url. This page is for engineers choosing a model or switching one: the graph and tools stay the same, only the model and provider arguments change.

Start here

Pick the page for the provider you use: OpenAI, Google or Anthropic. Each lists the install extra, the credentials it reads and the options specific to that backend. If you are not sure how a model string becomes a client, read Providers and adapters, and for the exact function signatures see the LLM utilities reference.

Provider choice is a small decision in 10xGraph. The same tools, state and checkpointer work with all of them, so you can start with one and swap later. What does not change is the production layer around the agent: the API server, auth and replay-safe tool calls described in Production runtime. If you want to see a first agent running before choosing, follow Your first agent.

Extras at a glance

The core package declares these install extras for providers: openai, google-genai, anthropic, anthropic-vertex and anthropic-bedrock. A separate realtime extra, built on google-genai, supports live audio sessions.

Available providers

create_llm_client and detect_provider recognise exactly three provider values:

Any other value raises ValueError: Unsupported provider.

When provider is omitted, detect_provider infers it from the model name: gemini-, imagen-, veo-, and chirp prefixes resolve to "google"; claude- and anthropic. resolve to "anthropic"; gpt-, o1-, o3-, and o4- resolve to "openai". A recognised provider/model prefix (for example gemini/gemini-2.5-flash or anthropic/claude-sonnet-4-5) selects the provider directly. Anything unrecognised falls back to "openai" and logs that it did so.

Other models

Self-hosted and third-party models served behind an OpenAI-compatible API, Ollama, vLLM, OpenRouter, and gateways of that shape, go through the OpenAI provider with a base_url.

agent = Agent(
    model="my-model",
    provider="openai",
    base_url="https://my-openai-compatible-gateway.example.com/v1",
    api_key="...",
)

Some of those only implement the legacy Chat Completions endpoint; pass api_style="chat" when the Responses API is not available.

Related docs

• Agents and tools
• LLM utilities reference