LLM Providers

🤖 Configure and optimize different LLM providers with ReasonKit.

Universal Compatibility: ReasonKit integrates seamlessly with Claude, Gemini, OpenAI, Cursor, VS Code, and any LLM provider. The same structured reasoning protocols work across all platforms, giving you flexibility without vendor lock-in.

ReasonKit supports multiple LLM providers, each with different strengths, pricing, and capabilities.

Supported Providers

Provider Configuration

Anthropic (Recommended)

Claude models provide the best reasoning quality for ThinkTools.

# Set API key
export ANTHROPIC_API_KEY="sk-ant-..."

# Use explicitly
rk think "question" --provider anthropic --model claude-sonnet-4-20250514

Config file:

[providers.anthropic]
api_key = "${ANTHROPIC_API_KEY}"  # Use env var
model = "claude-sonnet-4-20250514"
max_tokens = 4096

Available models:

OpenAI

export OPENAI_API_KEY="sk-..."

rk think "question" --provider openai --model gpt-4-turbo

Config file:

[providers.openai]
api_key = "${OPENAI_API_KEY}"
model = "gpt-4-turbo"
organization_id = "org-..."  # Optional
base_url = "https://api.openai.com/v1"  # For proxies

Available models:

OpenRouter

Access 300+ models through a single API. Great for cost optimization and experimentation.

export OPENROUTER_API_KEY="sk-or-..."

rk think "question" --provider openrouter --model anthropic/claude-sonnet-4

Config file:

[providers.openrouter]
api_key = "${OPENROUTER_API_KEY}"
model = "anthropic/claude-sonnet-4"
site_url = "https://yourapp.com"  # For rankings
site_name = "Your App"

Popular models:

Ollama (Local)

Run models locally for privacy and zero API costs.

# Start Ollama
ollama serve

# Pull a model
ollama pull llama3.2

# Use with ReasonKit
rk think "question" --provider ollama --model llama3.2

Config file:

[providers.ollama]
host = "http://localhost:11434"
model = "llama3.2"

Recommended models:

Google Gemini

export GOOGLE_API_KEY="..."

rk think "question" --provider google --model gemini-pro

Config file:

[providers.google]
api_key = "${GOOGLE_API_KEY}"
model = "gemini-pro"

Provider Selection

Automatic Selection

By default, ReasonKit auto-selects based on available API keys:

# Priority order:
# 1. ANTHROPIC_API_KEY
# 2. OPENAI_API_KEY
# 3. OPENROUTER_API_KEY
# 4. GOOGLE_API_KEY
# 5. Ollama (if running)

rk think "question"  # Uses first available

Per-Profile Provider

Configure different providers for different profiles:

[profiles.quick]
provider = "ollama"
model = "llama3.2"

[profiles.balanced]
provider = "anthropic"
model = "claude-sonnet-4-20250514"

[profiles.deep]
provider = "anthropic"
model = "claude-opus-4-20250514"

Cost Optimization

# Use cheaper models for simple tasks
[profiles.quick]
provider = "openrouter"
model = "mistralai/mistral-7b-instruct"  # Very cheap

[profiles.balanced]
provider = "openrouter"
model = "anthropic/claude-sonnet-4"  # Good balance

[profiles.paranoid]
provider = "anthropic"
model = "claude-opus-4-20250514"  # Best quality

Advanced Configuration

Timeouts

[providers.anthropic]
timeout_secs = 120
connect_timeout_secs = 10

Retries

[providers.anthropic]
max_retries = 3
retry_delay_ms = 1000
retry_multiplier = 2.0  # Exponential backoff

Rate Limiting

[providers.anthropic]
requests_per_minute = 50
tokens_per_minute = 100000

Custom Endpoints

For proxies or enterprise deployments:

[providers.openai]
base_url = "https://your-proxy.com/v1"
api_key = "${PROXY_API_KEY}"

Temperature and Sampling

[providers.anthropic]
temperature = 0.7        # 0.0-1.0, lower = more deterministic
top_p = 0.9             # Nucleus sampling
top_k = 40              # Top-k sampling

Provider-Specific Features

Anthropic Extended Thinking

Enable extended thinking for complex analysis:

[providers.anthropic]
extended_thinking = true
thinking_budget = 16000  # Max thinking tokens

OpenAI Function Calling

[providers.openai]
function_calling = true

OpenRouter Fallbacks

[providers.openrouter]
model = "anthropic/claude-sonnet-4"
fallback_models = [
    "openai/gpt-4-turbo",
    "google/gemini-pro",
]

Monitoring and Debugging

Token Usage

# Show token usage after each analysis
rk think "question" --verbose

# Output includes:
# Tokens: 1,234 prompt + 567 completion = 1,801 total
# Cost: ~$0.0054

Request Logging

# Log all API requests (for debugging)
export RK_DEBUG_API=true
rk think "question"

Provider Health Check

# Check if provider is working
rk provider test anthropic
rk provider test openai
rk provider test ollama

Switching Providers

Migration Checklist

When switching providers:

1. Test compatibility — Run same prompts, compare quality
2. Adjust timeouts — Different providers have different latencies
3. Check token limits — Models have different context windows
4. Update rate limits — Different quotas per provider
5. Review costs — Pricing varies significantly

Quality Comparison

# Run same analysis with different providers
rk think "question" --provider anthropic --output json > anthropic.json
rk think "question" --provider openai --output json > openai.json
rk think "question" --provider ollama --output json > ollama.json

# Compare results
diff anthropic.json openai.json

Troubleshooting

Common Issues

Error Codes

Related

• Configuration — General configuration
• Environment Variables — API key setup
• Architecture — Provider layer internals