Model setup and configuration

mini-swe-agent supports virtually every LLM provider through four backends: litellm (the default, covering OpenAI, Anthropic, Google, and dozens more), OpenRouter, Portkey, and Requesty. This page explains how to pick a provider, set your API key, and configure advanced options like reasoning effort and cost tracking.

Quick setup

The fastest way to get started is the interactive wizard, which sets your default model and API key in one step:

mini-extra config setup

This runs automatically the first time you launch mini. If you skipped it, run the command above. Your settings are saved to mini’s global config file and take effect for every subsequent run.

Manual configuration

You can also set or update individual values with mini-extra config set:

# Set the default model
mini-extra config set MSWEA_MODEL_NAME "anthropic/claude-opus-4-6-20260205"

# Set an API key
mini-extra config set ANTHROPIC_API_KEY "sk-ant-..."

Both values can also be exported as environment variables if you prefer not to use the config file:

export MSWEA_MODEL_NAME="anthropic/claude-opus-4-6-20260205"
export ANTHROPIC_API_KEY="sk-ant-..."

Always include the provider prefix in model names (e.g., anthropic/claude-..., openai/gpt-..., gemini/gemini-...). The prefix tells litellm which provider to route the request to.

Provider setup

Anthropic
OpenAI
Google (Gemini)
OpenRouter
Local models

Anthropic’s Claude models are the recommended choice for most tasks. mini-swe-agent automatically applies prompt caching for any model whose name contains anthropic, claude, sonnet, or opus.Recommended models:

Model	Use case
`anthropic/claude-opus-4-6-20260205`	Best overall quality; recommended for complex tasks
`anthropic/claude-sonnet-4-5-20250929`	Fast and cost-effective for everyday work
`anthropic/claude-haiku-4-5-20251001`	Fastest and cheapest; good for simple edits

Set your API key:

mini-extra config set ANTHROPIC_API_KEY "sk-ant-..."

Set the default model:

mini-extra config set MSWEA_MODEL_NAME "anthropic/claude-opus-4-6-20260205"

OpenAI’s GPT models are supported via litellm. For models that support the Responses API (like GPT-5), you can use the litellm_response model class for native tool calling.Recommended models:

Model	Use case
`openai/gpt-5.4`	Best OpenAI quality
`openai/gpt-5.4-mini`	Fast and affordable

Set your API key:

mini-extra config set OPENAI_API_KEY "sk-..."

Set the default model:

mini-extra config set MSWEA_MODEL_NAME "openai/gpt-5.4"

Optional: use the Responses API for GPT-5 with reasoning control:

# In your agent config file
model:
  model_class: "litellm_response"
  model_name: "openai/gpt-5-mini"
  model_kwargs:
    drop_params: true
    reasoning:
      effort: "high"

Google’s Gemini models are supported via litellm using the gemini/ prefix.Recommended model:

Model	Use case
`gemini/gemini-2.5-pro`	Strong coding performance; >74% on SWE-bench

Set your API key:

mini-extra config set GEMINI_API_KEY "..."

Set the default model:

mini-extra config set MSWEA_MODEL_NAME "gemini/gemini-2.5-pro"

OpenRouter provides a single endpoint for accessing many model providers. Use the openrouter model class to route requests directly through their API.Set your API key:

mini-extra config set OPENROUTER_API_KEY "sk-or-..."

Run with OpenRouter from the CLI:

mini -m "moonshotai/kimi-k2-0905" --model-class openrouter

Or configure in your agent config file:

model:
  model_name: "moonshotai/kimi-k2-0905"
  model_class: "openrouter"
  model_kwargs:
    temperature: 0.0
    provider:
      allow_fallbacks: false
      only: ["Moonshot AI"]

Locking provider.only is useful for benchmarking when you need consistent cost behavior. For everyday use, leave it unset to benefit from OpenRouter’s automatic fallback and routing.

Local models (Ollama, vLLM, LM Studio, etc.) are supported via litellm’s custom_llm_provider and api_base options. Configure them in your agent config file:

model:
  model_name: "my-local-model"
  model_kwargs:
    custom_llm_provider: "openai"
    api_base: "http://localhost:8000/v1"

Cost tracking for local modelslitellm will raise an error if it cannot find cost data for your model. For local/free models, disable cost tracking:

export MSWEA_COST_TRACKING="ignore_errors"

Or set it per-model in your config file:

model:
  model_name: "my-local-model"
  cost_tracking: "ignore_errors"
  model_kwargs:
    custom_llm_provider: "openai"
    api_base: "http://localhost:8000/v1"

To register a custom model for accurate cost tracking, set LITELLM_MODEL_REGISTRY_PATH to a JSON file mapping model names to their litellm configuration. See global configuration for details.

Environment variable reference

Variable	Description
`MSWEA_MODEL_NAME`	Default model used when no `-m` flag is passed to `mini`
`ANTHROPIC_API_KEY`	API key for Anthropic models
`OPENAI_API_KEY`	API key for OpenAI models
`GEMINI_API_KEY`	API key for Google Gemini models
`OPENROUTER_API_KEY`	API key for OpenRouter
`PORTKEY_API_KEY`	API key for Portkey
`PORTKEY_VIRTUAL_KEY`	Optional virtual key for Portkey routing
`MSWEA_COST_TRACKING`	Set to `ignore_errors` to suppress cost tracking errors (useful for local/free models)
`MSWEA_GLOBAL_COST_LIMIT`	Abort the agent if cumulative cost exceeds this amount in USD
`MSWEA_GLOBAL_CALL_LIMIT`	Abort the agent after this many model calls
`LITELLM_MODEL_REGISTRY_PATH`	Path to a JSON file that registers custom models with litellm

litellm supports dozens of additional API key variables (e.g., DEEPSEEK_API_KEY, GROQ_API_KEY, MISTRAL_API_KEY). Run mini-extra config setup to be guided to the right one for your provider, or check the litellm providers list.

Advanced: extra model settings

To configure options like temperature or reasoning effort, edit your agent config file (the path is printed when you run mini). Open it quickly with:

mini-extra config edit

Examples:

# Temperature (not supported by all models)
model:
  model_name: "anthropic/claude-sonnet-4-5-20250929"
  model_kwargs:
    temperature: 0.0

# GPT-5 with reasoning effort
model:
  model_name: "openai/gpt-5.4"
  model_kwargs:
    drop_params: true
    reasoning_effort: "high"
    verbosity: "medium"

# Portkey with a virtual key
model:
  model_name: "@openai/gpt-5.1"
  model_class: "portkey"
  model_kwargs:
    reasoning_effort: "medium"
    verbosity: "medium"

Get Started

Usage

Batch & Benchmarks

Advanced

Reference

Model setup and configuration

Quick setup

Manual configuration

Provider setup

Environment variable reference

Advanced: extra model settings

Build docs developers (and LLMs) love

Get Started

Usage

Batch & Benchmarks

Advanced

Reference

Documentation Index

​Quick setup

​Manual configuration

​Provider setup

​Environment variable reference

​Advanced: extra model settings

Build docs developers (and LLMs) love

Quick setup

Manual configuration

Provider setup

Environment variable reference

Advanced: extra model settings