@llms-sdk/bridge

Use OpenAI, Google, and other LLM providers with Claude Code by intercepting and transforming API requests.

Not that anything can beat Opus, Sonnet and a Claude Max plan. But you can try that fools errand now. Go get Claude Max.

Quick Start

# No installation required - use npx
npx @llms-sdk/bridge                     # Show available providers
npx @llms-sdk/bridge openai              # Show OpenAI models
npx @llms-sdk/bridge openai gpt-4o       # Run Claude Code with GPT-4

# Set API keys (optional - can specify per-command with --apiKey)
export OPENAI_API_KEY=sk-...
export GOOGLE_API_KEY=...

# Advanced usage
npx @llms-sdk/bridge openai gpt-4o --apiKey sk-...                           # Custom API key
npx @llms-sdk/bridge openai llama3.2 --baseURL http://localhost:11434/v1     # Local Ollama
npx @llms-sdk/bridge openai gpt-4o --baseURL https://openrouter.ai/api/v1 --apiKey sk-or-... # OpenRouter
npx @llms-sdk/bridge openai gpt-4o --debug                          # Enable debug logs
npx @llms-sdk/bridge --trace -p "Hello world"                       # Spy on Claude ↔ Anthropic communication

# All Claude Code arguments work
npx @llms-sdk/bridge google gemini-2.5-pro-preview-05-06 --resume --continue
npx @llms-sdk/bridge openai o4-mini -p "Hello world"

How It Works

Claude Code only works with Anthropic models. This tool intercepts Claude Code's API calls and routes them to other providers while preserving (almost) full tool compatibility.

1. Spawn Claude Code as subprocess with custom Node.js loader
2. Patch global fetch() to intercept api.anthropic.com/v1/messages requests
3. Transform Anthropic requests → unified LLMS SDK format → provider API
4. Stream provider responses back in Anthropic SSE format

Limitations

This is a glorified hack that pretends other models are Claude. Here's what breaks:

Completely Broken:

• 🚫 Token usage/cost reporting (Claude Code's displays will lie to you)
• 🚫 Image uploads (drag/drop, paste, file paths - Claude Code expects Anthropic's servers)
• 🚫 Input caching (Claude Code's prompt caching isn't implemented - enjoy higher costs!)
• 🚫 Web search/fetch tools (Anthropic-specific magic)

Somewhat Janky:

• 🤷 Model-specific features don't translate (Claude's artifacts, GPT's reasoning modes)
• 🤷 Thinking/reasoning output formatting differs between providers
• 🤷 Error messages might be cryptic (provider auth failures won't surface clearly)
• 🤷 Tool schemas get converted (JSON Schema ↔ Zod) - usually works, sometimes doesn't
• 🤷 Streaming behavior has subtle differences despite SSE format conversion

OpenAI Specific Rant: OpenAI, put the goddamn thinking tokens into your API responses, you cowards. We're all tired of your "reasoning effort" parameter nonsense.

There are definitely mystery bugs hiding in the corners. You've been warned. 🐛

Development

Setup:

git clone https://github.com/lennylmiller/llms-sdk
cd llms-sdk && npm install && npm run dev

This starts compilation in watch mode for all packages and apps. Code changes are reflected immediately. Use npx tsx src/cli.ts for on-the-fly compilation & testing.

Testing:

npm run test:all         # All tests
npm run test:unit        # Unit tests
npm run test:core        # CLI functionality
npm run test:tools       # Tool integration
npm run test:providers   # Multi-provider

Debugging:

# Enable debug logging
npx @llms-sdk/bridge openai gpt-4o --debug
cat .claude-bridge/requests-*.jsonl     # Raw request/response pairs
cat .claude-bridge/transformed-*.jsonl  # Transformation details
cat .claude-bridge/context-*.jsonl      # Message contexts and transform status

# Trace mode - spy on Claude Code ↔ Anthropic communication
npx @llms-sdk/bridge --trace -p "Test prompt"  # Normal Claude Code, but logs all requests/responses
cat .claude-bridge/trace-*.jsonl        # See system prompts, tools, thinking status, messages
cat .claude-bridge/requests-*.jsonl     # Raw request/response pairs

# VS Code debugging (requires patching Claude to disable anti-debugging)
npx tsx src/cli.ts <arguments> --patch-claude   # In JavaScript Debug Terminal

Bundling:

This package uses a hybrid bundling approach:

• Core bridge logic is bundled with LLMS SDK package
• LLM provider SDKs (@anthropic-ai/sdk, openai, @google/genai, etc.) remain external dependencies
• This avoids Node.js dynamic require issues while keeping dependencies manageable

Core Files:

• src/cli.ts - CLI with provider discovery
• src/interceptor.ts - Fetch interception & client creation
• src/transforms/ - Request/response transformations
• src/utils/ - SSE streaming, logging, parsing