Claude Code Proxy

TypeScript server that exposes an Anthropic-compatible API backed by Azure OpenAI, allowing Claude Code to communicate with Azure deployments via localhost:9999.

Prerequisites

• Node.js 18+
• ~/.codex/config.toml with an azure model_provider entry (the existing CLI config works)
• Environment variable matching the env_key defined in the config (e.g. OPENAI_API_KEY)

Install

npm install

Run

# optional: override port
export PORT=9999

npm run dev

Or run the compiled CLI directly:

npm run build
node dist/cli.js --port 9999

After publishing you can install it globally:

npm install -g openai-claude
openai-claude --port 9999 --host 0.0.0.0

Set Claude Code to use the proxy:

export ANTHROPIC_BASE_URL="http://0.0.0.0:9999"
export ANTHROPIC_AUTH_TOKEN="your-shared-secret"
export ANTHROPIC_MODEL="gpt-5-codex"
# optional: verbose proxy logging
# export DEBUG=true

By default the server requires x-api-key/Bearer auth that matches ANTHROPIC_AUTH_TOKEN. If you omit that environment variable the proxy skips authentication.

Notes

• The proxy reads ~/.codex/config.toml on startup to discover the Azure endpoint, wire API, and API key environment variable.
• If the configured env_key (e.g. OPENAI_API_KEY) is unset, the proxy falls back to ANTHROPIC_AUTH_TOKEN so you can reuse the same secret for both Claude Code and Azure.
• Optional model_reasoning_effort in ~/.codex/config.toml is forwarded to Azure via the reasoning.effort field unless a request provides its own override.
• /v1/messages now supports both standard and streaming responses. Streaming is proxied as SSE using Anthropic-compatible events.
• Tool/function calls are translated between Anthropic and Azure responses formats, so Claude Code can invoke tools and return results through the proxy.
• Requests are forwarded to the Azure responses API and the response is translated back to the Anthropic message format expected by Claude Code.
• Tool outputs should be returned to the assistant as Anthropic tool_result content blocks; the proxy relays them to Azure as function_call_output entries automatically.
• npm run build is executed automatically before npm publish so the generated dist/ folder is packaged; the published module exposes a global openai-claude executable.
• Image content blocks from Claude Code (type: "image") are converted to Azure input_image payloads, so vision prompts work through the proxy.