OpenAI Responses Anthropic Proxy

中文说明

Standalone local HTTP proxy that exposes Anthropic-compatible POST /v1/messages semantics to Claude Code while forwarding requests to OpenAI's Responses API or OpenAI-compatible upstreams.

What It Does

• accepts Anthropic-style POST /v1/messages
• forwards to POST /v1/responses
• supports streaming text responses
• translates OpenAI function calls into Anthropic tool_use
• preserves multi-turn tool continuation with previous_response_id
• works with OpenAI and compatible providers that implement the Responses API
• emits one structured log line per request

Runtime Requirement

• Bun >= 1.3

Quick Start

1. Install dependencies:

bun install

2. Create your env file:

cp .env.example .env

3. Start the proxy:

set -a
source .env
set +a
bun run start

You can also pass the same values as CLI flags:

bun run start -- \
  --listen-port 4141 \
  --upstream-url https://api.openai.com \
  --upstream-key "$OPENAI_API_KEY" \
  --upstream-model gpt-4.1 \
  --state-file .openai-responses-anthropic-proxy-state.json

Helper Scripts

The project includes lightweight lifecycle scripts for running the proxy as a background service from the project directory:

./start-proxy.sh
./proxy-status.sh
./stop-proxy.sh

What they do:

• start-proxy.sh starts the proxy in the background, writes the PID to .openai-responses-anthropic-proxy.pid, and appends logs to .openai-responses-anthropic-proxy.log
• proxy-status.sh reports whether the proxy is running, stale-pid, or stopped
• stop-proxy.sh stops the running proxy using the saved PID

If needed, you can override the PID file path with OPENAI_RESPONSES_PROXY_PID_FILE.

Claude Code Configuration

Point Claude Code to the local proxy:

export ANTHROPIC_BASE_URL=http://127.0.0.1:4141
export ANTHROPIC_API_KEY=dummy
export ANTHROPIC_MODEL=my-proxy-codex

dummy is only a placeholder required by Claude Code. The real upstream key is configured on the proxy side via .env or CLI flags. ANTHROPIC_MODEL is the name Claude Code will display locally; the proxy still forwards to the real upstream model configured with OPENAI_RESPONSES_UPSTREAM_MODEL or --upstream-model.

By default the proxy persists tool continuation state to .openai-responses-anthropic-proxy-state.json, so previous_response_id survives a restart. Override it with OPENAI_RESPONSES_STATE_FILE or --state-file.

Logs

Each POST /v1/messages request emits one structured line to stdout:

[openai-responses-proxy] request {"ts":"2026-04-06T03:06:30.090Z","method":"POST","path":"/v1/messages","model":"claude-sonnet-test","stream":true,"status":200,"upstream_status":200,"duration_ms":1209,"error":null}

This lets you confirm that Claude Code is actually going through the proxy.

Verified Usage

The screenshot below shows a real Claude Code session using this proxy to ask 分析这个项目的结构 and receive a normal structured answer instead of a tool continuation error:

Change Log

• 2026-04-08
  • added richer Responses → Anthropic translation for reasoning summaries, refusal text, cache-token usage fields, and incomplete/max-token stop reason
  • expanded streaming SSE handling for reasoning and refusal deltas
  • added regression coverage for reasoning, refusal, usage normalization, reasoning effort mapping, and streaming behavior
  • added start-proxy.sh, stop-proxy.sh, and proxy-status.sh for local proxy lifecycle management
• 2026-04-07
  • added verified usage screenshot to the README
  • fixed tool continuation compatibility for Responses upstreams
• 2026-04-06
  • initial standalone proxy release
  • added request logging with local timestamps and timezone offset

Environment Variables

• OPENAI_RESPONSES_PROXY_HOST
• OPENAI_RESPONSES_PROXY_PORT
• OPENAI_RESPONSES_PROXY_PID_FILE
• OPENAI_RESPONSES_UPSTREAM_URL
• OPENAI_RESPONSES_UPSTREAM_KEY
• OPENAI_RESPONSES_UPSTREAM_MODEL
• OPENAI_RESPONSES_STATE_FILE

Docker

Build the image:

docker build -t openai-responses-anthropic-proxy .

Run it:

docker run --rm \
  -p 4141:4141 \
  -e OPENAI_RESPONSES_PROXY_HOST=0.0.0.0 \
  -e OPENAI_RESPONSES_PROXY_PORT=4141 \
  -e OPENAI_RESPONSES_UPSTREAM_URL=https://api.openai.com \
  -e OPENAI_RESPONSES_UPSTREAM_KEY="$OPENAI_API_KEY" \
  -e OPENAI_RESPONSES_UPSTREAM_MODEL=gpt-4.1 \
  openai-responses-anthropic-proxy

Build A Single Binary

bun run build:binary

This writes a standalone executable to:

dist/openai-responses-anthropic-proxy

Development Commands

bun run test
bun run typecheck

Project Structure

src/cli.ts
src/server.ts
src/translate.ts
src/state.ts
src/config.ts
src/types.ts
start-proxy.sh
stop-proxy.sh
proxy-status.sh