Codex WebSocket-to-HTTPS Proxy

A lightweight local proxy for Codex model requests. Codex can keep its ChatGPT login and remote-control behavior while model requests are billed through an OpenAI-compatible upstream relay.

Build

go mod tidy
go build -o codex-ws-proxy .

Configure

Place a.json next to the binary, or in the working directory when running with go run ..

{
  "listen": "127.0.0.1:39493",
  "upstream_base_url": "https://example-relay.com/v1",
  "local_base_path": "/v1",
  "api_key": "replace-with-api-key",
  "auth_header": "Authorization",
  "auth_scheme": "Bearer",
  "websocket_mode": "bridge",
  "websocket_compression": false,
  "timeout_seconds": 600,
  "log_requests": false
}

Point Codex at the proxy:

openai_base_url = "http://127.0.0.1:39493/v1"

Behavior

• Normal HTTP requests under /v1 are forwarded to the upstream base URL.
• WebSocket requests to /v1/responses are accepted locally and converted to HTTP POST /v1/responses with Accept: text/event-stream.
• Incoming Codex, ChatGPT, OAuth, cookie, and OpenAI-specific headers are stripped before forwarding.
• Upstream auth is injected from a.json.
• Codex WebSocket-only fields type, generate, and previous_response_id are not sent to HTTP upstream.
• Prewarm frames with generate:false are answered locally.
• Per-connection in-memory state reconstructs tool-call context for stateless upstream relays.
• Adapted /v1/models responses advertise the Codex fast speed tier and preserve service_tier on model requests.
• WebSocket compression is disabled for Codex client compatibility.

Verify

go test ./...