Models

Multi-provider LLM support and model configuration

Klisk supports OpenAI models natively and any LiteLLM-compatible provider, giving you access to hundreds of language models from different providers.

Model string format

The model parameter in define_agent() uses a convention to determine whether to use OpenAI directly or route through LiteLLM:

OpenAI (no slash) · string

No / character → native OpenAI model

define_agent(name="Agent", model="gpt-5.2")         # OpenAI native
define_agent(name="Agent", model="gpt-4.1")        # OpenAI native
define_agent(name="Agent", model="o3")             # OpenAI native

OpenAI (with prefix) · string

openai/ prefix → strips prefix, uses native OpenAI

define_agent(name="Agent", model="openai/gpt-5.2")  # OpenAI native (prefix stripped)

LiteLLM (provider prefix) · string

provider/ format → routes through LiteLLM

define_agent(name="Agent", model="anthropic/claude-sonnet-4-5-20250929")
define_agent(name="Agent", model="gemini/gemini-2.5-flash")
define_agent(name="Agent", model="mistral/mistral-large-latest")

OpenAI models

OpenAI models work out of the box. Set your API key:

.env

OPENAI_API_KEY=sk-...

Then reference the model directly:

from klisk import define_agent

agent = define_agent(
    name="Assistant",
    model="gpt-5.2",  # Default if omitted
)

Available OpenAI models:

• gpt-5.2, gpt-5.1 — Latest GPT-5 series
• gpt-4.1, gpt-4o, gpt-4o-mini — GPT-4 series
• o3, o4-mini, o1 — Reasoning models

The default model is gpt-5.2 if you don’t specify the model parameter.

LiteLLM providers

For non-OpenAI models, Klisk uses LiteLLM to provide a unified interface to 100+ LLM providers.

Installation

LiteLLM is automatically installed when you use a non-OpenAI model string:

# First time this runs, Klisk will prompt to install litellm
agent = define_agent(
    name="Agent",
    model="anthropic/claude-sonnet-4-5-20250929",
)

Or install it manually:

pip install litellm

API keys

Klisk auto-detects API keys from environment variables based on the provider prefix:

.env

ANTHROPIC_API_KEY=sk-ant-...
GEMINI_API_KEY=...
MISTRAL_API_KEY=...
COHERE_API_KEY=...

The pattern is {PROVIDER}_API_KEY where PROVIDER is the uppercase version of the prefix:

Model String	Environment Variable
anthropic/...	ANTHROPIC_API_KEY
gemini/...	GEMINI_API_KEY
mistral/...	MISTRAL_API_KEY
cohere/...	COHERE_API_KEY
azure/...	AZURE_API_KEY

If you’re using a non-OpenAI model and OPENAI_API_KEY is not set, Klisk automatically disables OpenAI tracing to prevent errors.

Model resolution

When you specify a LiteLLM model, Klisk:

1. Enables LiteLLM serializer compatibility patch to suppress Pydantic warnings
2. Imports LitellmModel from the OpenAI Agents SDK
3. Extracts the provider prefix (e.g., anthropic from anthropic/claude-sonnet-4-5)
4. Looks up {PROVIDER}_API_KEY from environment variables
5. Creates a LitellmModel(model=model_str, api_key=api_key) instance

This happens automatically in define_agent() via the _resolve_model() function in primitives.py:19.

Popular providers

Anthropic (Claude)

define_agent(
    name="Agent",
    model="anthropic/claude-sonnet-4-5-20250929",
)

Available models:

• claude-sonnet-4-5-20250929 — Latest Claude Sonnet
• claude-opus-4 — Most capable
• claude-haiku-4 — Fast and efficient

Google (Gemini)

define_agent(
    name="Agent",
    model="gemini/gemini-2.5-flash",
)

Available models:

• gemini-2.5-flash — Fast, cost-effective
• gemini-2.5-pro — Advanced reasoning
• gemini-1.5-pro — Previous generation

Mistral

define_agent(
    name="Agent",
    model="mistral/mistral-large-latest",
)

Available models:

• mistral-large-latest — Most capable
• mistral-medium-latest — Balanced
• mistral-small-latest — Fast

Azure OpenAI

define_agent(
    name="Agent",
    model="azure/gpt-4o",
)

Requires additional Azure-specific environment variables. See LiteLLM Azure docs.

Reasoning effort

For reasoning models (o-series and gpt-5+), you can control how much reasoning the model applies:

define_agent(
    name="Analyst",
    model="o3",
    reasoning_effort="high",
)

Valid values: "none", "minimal", "low", "medium", "high", "xhigh"

LiteLLM automatically translates reasoning_effort to each provider’s equivalent parameter. For example, Anthropic’s thinking parameter.

Only set reasoning_effort for reasoning models (o1, o3, o4-mini, gpt-5.1, gpt-5.2). It has no effect on standard models like gpt-4o or claude-sonnet-4-5.

When to use reasoning effort

Value	Use Case	Cost
"none" or "minimal"	Simple, routine tasks	Lower
"low" or "medium"	Standard problem-solving	Moderate
"high" or "xhigh"	Complex analysis, math, code	Higher

# Simple data extraction
define_agent(
    name="Extractor",
    model="gpt-5.2",
    reasoning_effort="minimal",
)

# Complex mathematical analysis
define_agent(
    name="MathSolver",
    model="o3",
    reasoning_effort="xhigh",
)

Temperature

Control response randomness with the temperature parameter:

define_agent(
    name="CreativeWriter",
    model="gpt-5.2",
    temperature=1.5,  # More creative, diverse outputs
)

define_agent(
    name="DataAnalyst",
    model="gpt-5.2",
    temperature=0.2,  # More focused, deterministic
)

Range: 0.0 (deterministic) to 2.0 (very random)

Defaults: Each model has its own default (usually around 0.7 to 1.0)

Builtin tools compatibility

Builtin tools (web_search, code_interpreter, file_search, image_generation) only work with OpenAI models:

# ✅ Works
define_agent(
    name="Researcher",
    model="gpt-5.2",
    builtin_tools=["web_search"],
)

# ❌ Warns and disables builtin tools
define_agent(
    name="Researcher",
    model="anthropic/claude-sonnet-4-5-20250929",
    builtin_tools=["web_search"],  # Will be disabled with a warning
)

This is checked at agent creation time in primitives.py:104. If you specify builtin tools with a non-OpenAI model, you’ll see:

Warning: Builtin tools (web_search) are only supported with OpenAI models
and will be disabled for 'anthropic/claude-sonnet-4-5-20250929'.

Complete example

from klisk import define_agent, get_tools

# OpenAI with reasoning
analyst = define_agent(
    name="DataAnalyst",
    model="o3",
    reasoning_effort="high",
    temperature=0.3,
    tools=get_tools("analyze"),
)

# Anthropic with custom tools
researcher = define_agent(
    name="Researcher",
    model="anthropic/claude-sonnet-4-5-20250929",
    temperature=0.7,
    tools=get_tools("search", "summarize"),
)

# Google with high creativity
writer = define_agent(
    name="Writer",
    model="gemini/gemini-2.5-flash",
    temperature=1.5,
    tools=get_tools("draft", "revise"),
)

# OpenAI with builtin tools
assistant = define_agent(
    name="Assistant",
    model="gpt-5.2",
    builtin_tools=["web_search", "code_interpreter"],
)

Next steps

Agents

Learn more about agent configuration

Builtin Tools

Explore provider-hosted tools for OpenAI models