LLM Providers

Git-Iris supports OpenAI, Anthropic, Google, OpenRouter, and Fireworks.

Provider Overview

Provider	Default Model	Fast Model	Context Window	API Key Env
OpenAI	gpt-6-astra	gpt-5.6-luna	1.05M	OPENAI_API_KEY
Anthropic	claude-opus-5	claude-haiku-4-5-20251001	1M	ANTHROPIC_API_KEY
Google	gemini-3.8-flash	gemini-3.5-flash-lite	1M	GOOGLE_API_KEY

OpenRouter defaults to anthropic/claude-opus-5 and uses OPENROUTER_API_KEY. Fireworks defaults to accounts/fireworks/models/deepseek-v4-pro-0813 and uses FIREWORKS_API_KEY.

Configuration Format

Each provider has its own section under [providers]:

[providers.PROVIDER_NAME]
api_key = "YOUR_API_KEY"
model = "model-name"           # Optional: primary model
fast_model = "fast-model-name" # Optional: for status updates
subagent_model = "worker-model-name" # Optional: defaults to the primary model
token_limit = 1050000          # Optional: context-window metadata

OpenAI Configuration

[providers.openai]
api_key = "sk-..."
model = "gpt-6-astra"
fast_model = "gpt-5.6-luna"

CLI Setup

git-iris config --provider openai --api-key YOUR_API_KEY
git-iris config --provider openai --model gpt-6-astra

Environment Variable

export OPENAI_API_KEY="sk-..."

Anthropic Configuration

[providers.anthropic]
api_key = "sk-ant-..."
model = "claude-opus-5"
fast_model = "claude-haiku-4-5-20251001"

CLI Setup

git-iris config --provider anthropic --api-key YOUR_API_KEY
git-iris config --provider anthropic --model claude-opus-5

Environment Variable

export ANTHROPIC_API_KEY="sk-ant-..."

Legacy Aliases

The provider names claude and gemini are still supported as aliases for anthropic and google.

Google Configuration

[providers.google]
api_key = "your-google-api-key"
model = "gemini-3.8-flash"
fast_model = "gemini-3.5-flash-lite"

CLI Setup

git-iris config --provider google --api-key YOUR_API_KEY
git-iris config --provider google --model gemini-3.8-flash

Environment Variable

export GOOGLE_API_KEY="..."

OpenRouter Configuration

OpenRouter routes requests to hosted models using its own model IDs and API key. Iris uses the native OpenRouter integration to retain reasoning details across tool calls.

export OPENROUTER_API_KEY="sk-or-..."
git-iris config --provider openrouter --model anthropic/claude-opus-5

[providers.openrouter]
model = "anthropic/claude-opus-5"
fast_model = "anthropic/claude-haiku-4.5"

The Opus default uses high effort for main tasks and low effort for subagents. Use model IDs from OpenRouter's catalog and choose models with tool-calling support.

Fireworks Configuration

Fireworks uses its OpenAI-compatible Chat Completions endpoint. Model IDs include the account path.

export FIREWORKS_API_KEY="..."
git-iris config --provider fireworks --model accounts/fireworks/models/deepseek-v4-pro-0813

[providers.fireworks]
model = "accounts/fireworks/models/deepseek-v4-pro-0813"
fast_model = "accounts/fireworks/models/deepseek-v4-flash-0731"

For the default DeepSeek V4 models, Iris uses high effort for analysis and disables thinking for status messages. Fireworks promotes low and medium effort to high for this model family. Custom models retain their provider defaults unless you supply parameters. See Fireworks' model guide for serverless availability and supported features. An account-specific deployment can use its own model ID.

Switching Providers

Set Default Provider

git-iris config --provider anthropic

Override Per-Command

git-iris gen --provider openai
git-iris review --provider google

Additional Parameters

Provider-specific parameters can be set using --param:

git-iris config --provider openai --param reasoning='{"effort":"medium"}'
git-iris config --provider openai --param text='{"verbosity":"low"}'

Git-Iris parses valid JSON values here, so nested provider options work without extra config files. The --token-limit setting records context-window metadata; it does not change output budgets. Iris currently requests 16,384 output tokens for main tasks and 4,096 for subagents.

OpenAI defaults use the Responses API and choose reasoning by workflow:

• Main agent generations use reasoning = {"effort":"medium"}
• Subagents and parallel_analyze use reasoning = {"effort":"low"}
• Fast status messages use Luna with reasoning = {"effort":"none"}

Set reasoning yourself only when you want to override those defaults for every OpenAI request using that provider config.

In TOML:

[providers.openai]
api_key = "sk-..."

  [providers.openai.additional_params]
  reasoning = '{"effort":"medium"}'
  text = '{"verbosity":"low"}'

Token Limits

The model context window and output-token budget are separate limits. See Model Selection for role defaults and context guidance.

API Key Priority

API keys are loaded in this order:

1. Config file (~/.config/git-iris/config.toml)
2. Environment variable (OPENAI_API_KEY, etc.)

Project configs (.irisconfig) never contain API keys for security.

Verification

Check your provider configuration:

# View current configuration
cat ~/.config/git-iris/config.toml

# Test with a command
git-iris gen --print

If authentication fails, Git-Iris will tell you which environment variable to set.

Security Best Practices

• Never commit config.toml with API keys
• Use environment variables in CI/CD
• Restrict file permissions:

  chmod 600 ~/.config/git-iris/config.toml

• Rotate API keys periodically
• Use separate keys for different projects if needed