codex-cli-proxy 0.0.1

Local OpenAI and Anthropic-compatible chat proxy backed by the Codex CLI

Codex CLI Proxy

Overview

Codex CLI Proxy exposes a signed-in local Codex CLI as a localhost-only REST service. It accepts OpenAI Chat Completions and Anthropic Messages request formats and returns compatible response shapes.

Each request runs in its own empty temporary workspace, which is removed when the Codex turn finishes. The proxy therefore does not provide its own project directory as chat context. It supports text and base64-encoded image inputs.

Installation

Prerequisites:

• Python 3.12 or later
• uv
• A signed-in local codex CLI

Install the project dependencies:

uv sync

Run

uv run codex-cli-proxy

The server listens on http://127.0.0.1:9000 by default.

Endpoints

Method	Path	Description
GET	/healthz, /readyz	Proxy and Codex executable health status.
GET	/v1/models	Models available through the signed-in Codex CLI.
POST	/v1/chat/completions	OpenAI Chat Completions-compatible chat endpoint.
POST	/v1/messages	Anthropic Messages-compatible chat endpoint.

Both chat endpoints support stream: true; output is returned as compatibility streaming after the Codex turn completes.

Inputs

• Text content is supported.
• Images are supported as base64 payloads:
  • OpenAI: image_url or input_image with a data:image/...;base64,... URL.
  • Anthropic: image with a base64 source block.
  • JPEG, PNG, GIF, and WebP are accepted; up to 10 images and 20 MB per image.
• Audio and caller-supplied tools/functions are not supported by the local Codex CLI execution path and return a 400 error.

Use the standard model field to select a Codex model. If omitted, Codex uses its configured default. mode is accepted as a compatibility alias.

Configuration

Variable	Default	Purpose
CODEX_BIN	codex	Path or command for the local Codex CLI.
CODEX_SANDBOX	read-only	Codex sandbox mode for each request.
CODEX_TIMEOUT_SECONDS	900	Maximum duration of one Codex turn.
CODEX_MAX_CONCURRENCY	1	Maximum concurrent Codex turns.
HOST / PORT	127.0.0.1 / 9000	Server bind address and port.

Keep the service behind a trusted network boundary.

Examples

curl http://127.0.0.1:9000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"What is the capital of France?"}]}'

References

• Request examples (test.rest)
• OpenAI Chat Completions API
• Anthropic Messages API