Providers

pureagents supports multiple LLM providers.

Available providers

Provider	Default model	Environment variable
mistral	mistral-large-latest	MISTRAL_API_KEY
openai	gpt-5.6-luna	OPENAI_API_KEY
anthropic	claude-sonnet-5	ANTHROPIC_API_KEY

Mistral (default)

from pure_agents import Agent

# Uses MISTRAL_API_KEY from environment
agent = Agent()

# Or explicit
agent = Agent(api_key="your-mistral-key")

# Specific model
agent = Agent(model="mistral-small-latest")

OpenAI

agent = Agent(provider="openai")

# GPT-5.6 tiers, cheapest first
agent = Agent(provider="openai", model="gpt-5.6-luna")  # Default
agent = Agent(provider="openai", model="gpt-5.6-terra")  # Balanced
agent = Agent(provider="openai", model="gpt-5.6-sol")  # Highest capability

Anthropic

agent = Agent(provider="anthropic")

# Claude tiers, cheapest first
agent = Agent(provider="anthropic", model="claude-haiku-4-5")  # Fastest
agent = Agent(provider="anthropic", model="claude-sonnet-5")  # Default
agent = Agent(provider="anthropic", model="claude-opus-5")  # Agentic coding
agent = Agent(provider="anthropic", model="claude-fable-5")  # Long-running agents

Any OpenAI-compatible endpoint

Most of the ecosystem speaks the OpenAI wire format. Point base_url at it and you are done:

agent = Agent(base_url="http://localhost:11434/v1", model="llama3.3")

No API key is needed when the server does not check one. Pass api_key= when it does.

Common endpoints

Runtime	base_url	Key
Ollama	http://localhost:11434/v1	none
LM Studio	http://localhost:1234/v1	none
vLLM	http://localhost:8000/v1	none
OpenRouter	https://openrouter.ai/api/v1	OPENROUTER_API_KEY
Groq	https://api.groq.com/openai/v1	GROQ_API_KEY
Together	https://api.together.xyz/v1	TOGETHER_API_KEY

Model names are yours to supply. pureagents does not ship defaults for these because their catalogues move faster than a release cycle.

Registering one by name

If you use an endpoint more than once, give it a name:

from pure_agents import Agent, register_provider

register_provider("ollama", base_url="http://localhost:11434/v1")

agent = Agent(provider="ollama", model="llama3.3")

register_provider takes:

Parameter	Default	Description
base_url	Required	Endpoint root, without a trailing slash
env_var	None	Environment variable holding the key. Omit for a keyless local server
default_model	None	Used when the caller omits model=. Omit and callers must be explicit
dialect	"openai"	"openai" or "anthropic" wire format
headers	None	Extra headers every request needs
overwrite	False	Allow replacing an existing name

Built-in providers are protected: registering over openai, mistral or anthropic needs overwrite=True, and unregister_provider refuses them outright.

Extra headers

Some gateways want attribution headers:

agent = Agent(
    base_url="https://openrouter.ai/api/v1",
    model="anthropic/claude-sonnet-5",
    headers={"HTTP-Referer": "https://myapp.dev", "X-Title": "My App"},
)

Headers passed to Agent win over headers set on the registered provider.

Local models for development

Running the loop against a local model costs nothing, which makes it a good default while you iterate on prompts and tools:

import os

if os.environ.get("DEV"):
    agent = Agent(base_url="http://localhost:11434/v1", model="llama3.3")
else:
    agent = Agent(provider="anthropic")

Tool calling and streaming both work against Ollama, LM Studio and vLLM.

Explicit API key

You can pass the API key directly:

agent = Agent(provider="openai", api_key="sk-...")

warning

Don't commit API keys to your repository. Use environment variables.

Custom base URL

For self-hosted or proxy endpoints, modify the client directly:

from pure_agents.clients import LLMClient

client = LLMClient(api_key="your-key", base_url="https://your-proxy.com/v1")