sub2api_simple is a lightweight standalone gateway that exposes a unified API for multiple upstream AI accounts (Anthropic and OpenAI), with account pooling, failover, sticky sessions, and streaming support.

• Unified HTTP API for:
  • POST /v1/messages (Anthropic-compatible)
  • POST /v1/responses and POST /responses (OpenAI-compatible)
  • POST /v1/chat/completions (OpenAI Chat Completions-compatible)
  • GET /v1/models
  • GET /health
• Multi-account pool with:
  • Priority-based scheduling
  • Concurrency limits per account
  • Sticky session routing
  • Automatic failover on transient upstream errors
• OAuth and API key support:
  • OpenAI OAuth browser login flow (/auth/login)
  • OpenAI refresh token auto-refresh and persistence
  • Anthropic/OpenAI API key mode
• Streaming and non-streaming handling:
  • SSE passthrough for streaming requests
  • SSE aggregation for non-streaming clients when needed

• Go 1.21+

1. Copy and edit config:

cp config.example.json config.json

2. Build:

• Windows (PowerShell):

./build.ps1

• Linux/macOS:

bash ./build.sh

Default output is output/.

3. Run:

go run . -config config.json

or run the built binary:

• Windows:

./output/sub2api_simple.exe -config config.json

• Linux/macOS:

./output/sub2api_simple -config config.json

Client requests must provide one of the configured auth_tokens:

• Authorization: Bearer <token>
• or x-api-key: <token>

If an OpenAI OAuth account has no available token, the gateway logs a login URL on startup.

You can also open it manually:

http://localhost:8080/auth/login?account=<account-name>

The callback listener uses:

http://localhost:1455/auth/callback

See config.example.json for full options.

Main fields:

• listen_addr: bind address (for example :8080)
• auth_tokens: client-facing API tokens for this gateway
• accounts: upstream account definitions
• openai_default_instructions: default instructions by model (supports *)
• model_extra_params: default extra params by model (supports *)
• enable_request_log: enable request/response error logging
• enable_stream_debug_log: verbose stream line logging
• enable_model_debug_log: log upstream model request payloads and downstream client responses
• max_account_switches: maximum failover switches per request
• sticky_session_ttl: sticky routing TTL (duration string)
• stream_read_timeout: stream inactivity timeout (duration string)

Each account supports:

• platform: anthropic or openai
• type: oauth or api_key

Common auth patterns:

• Anthropic/OpenAI API key:
  • set type: "api_key" + api_key
• OpenAI OAuth refresh:
  • set type: "oauth" + refresh_token
• OpenAI OAuth browser login:
  • set type: "oauth" without token, then use /auth/login

• POST /v1/messages
  • Anthropic-compatible message API proxy
• POST /v1/responses
  • OpenAI-compatible responses API proxy
• POST /responses
  • Alias of /v1/responses
• POST /v1/chat/completions
  • OpenAI Chat Completions-compatible proxy (request/response format conversion)
• GET /v1/models
  • Combined model list based on configured platforms
• GET /health
  • Service and account runtime status
• GET /auth/login
  • Start OpenAI OAuth browser flow

go test ./... -run ^$

client_test.go expects a running gateway at http://localhost:8080:

go test ./...

.
├── main.go
├── internal/gateway
│   ├── oauth/       # token refresh, PKCE, OAuth session
│   ├── forwarder/   # failover policy and stream helpers
│   ├── router/      # endpoint dispatch
│   ├── claude/      # Claude request shaping
│   └── openai/      # OpenAI request shaping
├── cmd/chatclient   # optional interactive client
├── build.ps1
├── build.sh
└── config.example.json

This project is licensed under the MIT License. See LICENSE.