Providers

Agentix supports five LLM backends. All provider SDKs are bundled in the core install — no extra pip install is needed.

Supported providers

provider	Models	Default model
"anthropic"	Claude 3.x / 4.x	claude-sonnet-4-20250514
"openai"	GPT-4o, o-series, etc.	gpt-4o
"openai_compatible"	Any OpenAI API–compatible model	gpt-oss-120b
"gemini"	Gemini 2.x Flash/Pro	gemini-2.5-flash
"deepseek"	DeepSeek Chat / Coder	deepseek-chat

---

Anthropic

from agentix import AgentixAgentOptions

options = AgentixAgentOptions(
    provider="anthropic",
    model="claude-sonnet-4-20250514",
)

API key: AGENTIX_API_KEY

Extended thinking (effort)

Claude models support extended reasoning. Use effort as a shorthand or configure thinking directly:

# Shorthand — maps to thinking budget tokens
options = AgentixAgentOptions(
    provider="anthropic",
    effort="high",   # low=2k | medium=5k | high=10k | max=32k tokens
)

# Explicit
options = AgentixAgentOptions(
    provider="anthropic",
    thinking={"enabled": True, "budget_tokens": 10000},
)

ThinkingBlock entries appear in AssistantMessage.content when thinking is active.

Beta features

options = AgentixAgentOptions(
    provider="anthropic",
    betas=["interleaved-thinking-2025-05-14"],
)

Key rotation

For high-throughput use cases, supply a callback that returns a fresh API key before each LLM call:

import itertools

keys = itertools.cycle(["sk-ant-key1", "sk-ant-key2", "sk-ant-key3"])

options = AgentixAgentOptions(
    provider="anthropic",
    key_provider=lambda: next(keys),
)

---

OpenAI

options = AgentixAgentOptions(
    provider="openai",
    model="gpt-4o",
)

API key: AGENTIX_API_KEY

o-series reasoning models

options = AgentixAgentOptions(
    provider="openai",
    model="o3",
    effort="high",   # maps to reasoning_effort="high" for o-series
)

---

OpenAI-compatible

Use this provider for any endpoint that speaks the OpenAI API — Ollama, vLLM, LiteLLM, Cerebras, Together AI, Groq, and others.

base_url is required and must be set either in llm_options or via AGENTIX_BASE_URL.

# Ollama (local)
options = AgentixAgentOptions(
    provider="openai_compatible",
    model="llama3.2",
    llm_options={
        "base_url": "http://localhost:11434/v1",
        "api_key": "ollama",          # placeholder; Ollama ignores this
    },
)

# vLLM (hosted)
options = AgentixAgentOptions(
    provider="openai_compatible",
    model="meta-llama/Meta-Llama-3-70B-Instruct",
    llm_options={
        "base_url": "https://my-vllm-host/v1",
        "api_key": "...",
    },
)

Via environment variables:

export AGENTIX_PROVIDER=openai_compatible
export AGENTIX_BASE_URL=http://localhost:11434/v1
export AGENTIX_MODEL=llama3.2

AGENTIX_BASE_URL is only respected by openai_compatible and deepseek providers. It is ignored for anthropic, openai, and gemini.

---

Google Gemini

options = AgentixAgentOptions(
    provider="gemini",
    model="gemini-2.5-flash",
)

API key: AGENTIX_API_KEY

---

DeepSeek

options = AgentixAgentOptions(
    provider="deepseek",
    model="deepseek-chat",
)

API key: AGENTIX_API_KEY

The default base_url points to DeepSeek's official API. Override it to use a self-hosted or proxy endpoint:

options = AgentixAgentOptions(
    provider="deepseek",
    llm_options={"base_url": "https://my-proxy/v1"},
)

---

Selecting a provider via environment

Set the provider and model without touching code — useful for deploying the same agent to different backends:

export AGENTIX_PROVIDER=openai
export AGENTIX_MODEL=gpt-4o
export AGENTIX_API_KEY=sk-...

Then use a minimal options object:

options = AgentixAgentOptions(name="my-agent")
# provider="openai", model="gpt-4o" filled from environment

Environment variables are only applied when the corresponding field hasn't been explicitly set in code. Explicit fields always win.

---

Fallback model

Configure a fallback for when the primary model fails after all retries:

options = AgentixAgentOptions(
    provider="anthropic",
    model="claude-opus-4-20250514",
    fallback_model="claude-sonnet-4-20250514",  # used after retry exhaustion
)

---

Inline API key

Pass an API key directly to avoid relying on environment variables. Agentix redacts keys in logs and __repr__ automatically:

options = AgentixAgentOptions(
    provider="anthropic",
    llm_options={"api_key": "sk-ant-..."},
)

Prefer environment variables in production. Inline keys are convenient for short scripts but easy to accidentally commit or log.