Codex-Rosetta

English Version | 中文版

Codex-Rosetta is an LLM gateway based on LLM-Rosetta. It focuses on connecting third-party LLM APIs to Codex while improving tool calling and plugin behavior.

Fork Focus

This project is forked from Oaklight/llm-rosetta. This fork focuses on converting Chat Completions-compatible APIs to the Responses API, adapting tool-call semantics so open models work better in Codex, and aggregating multiple providers behind one gateway. The agent-facing generation API only exposes OpenAI Responses; Chat Completions, Anthropic Messages, and Google GenAI formats are retained as upstream target formats, not downstream client surfaces.

Installation

Clone the repository, enter it, and install the package:

git clone https://github.com/iBobbyTS/codex-rosetta.git
cd codex-rosetta
python -m pip install -U '.[gateway]'

The gateway extra installs the audited AEAD dependency used for exact, restart-safe tool-history persistence. Core conversion-only consumers may still install . without gateway dependencies.

Usage

Initialize the gateway configuration once:

codex-rosetta-gateway init

Initialization generates a mandatory Admin password and gateway access key in the owner-only config file. Store both securely; protected /v1 requests must send the generated access key as a Bearer token. See Gateway security and authentication.

Start the local gateway each time you use it:

codex-rosetta-gateway --host 127.0.0.1 --log-level warning

The default warning level shows only warnings and errors. Use --log-level stats to maintain request counts by original upstream model name on one terminal line, info to print request summaries, or error to show only errors. In stats mode, warnings and errors start on a new line; the counters resume on the next request. Use the WebUI Request Log for complete request history and Gateway Logs for streaming trace diagnostics.

Full Documentation

• English user documentation
• Gateway security and authentication
• Developer documentation

Problems This Project Addresses

Using third-party models in Codex usually runs into several issues:

• Providers may only expose a Chat Completions API.
• Models may not know how to edit files with apply_patch, so they fall back to sed, Python scripts, or other shell commands.
• Built-in Codex flows such as Goal and subagents may behave incorrectly.
• Models may not proactively call plugins.
• Some models do not support multimodal image understanding.
• Computer use and browser use may be unreliable.
• Models may not match the intended reasoning depth.

This project aims to improve those behaviors so strong models such as DeepSeek V4 Pro, GLM-5.x, and Qwen3.7 can run smoothly in Codex, with lower cost while still using Codex's advanced agent capabilities.

Currently solved, but not yet heavily production-tested:

• Responses API conversion: Oaklight/llm-rosetta provides the project base, core protocol conversion, and a simple web UI.
• Code editing tool translation: because these models often recommend Claude Code as their preferred coding agent, this project references Claude Code-style tool definitions. Models can emit familiar tool calls, and Rosetta converts them back to apply_patch or other Codex-native calls.
• Input-cache preservation: because the gateway intercepts and rewrites tool calls, the provider-side cache and Codex's local session history can otherwise diverge. Rosetta rewrites historical tool calls in outgoing requests so provider input caches can still match.
• Goal, TODO, Plan, and Subagent flows have been tested successfully.
• Work-process folding, with the trade-off that streaming is lost, but the behavior can now be toggled on or off.
• Model reasoning-depth mapping.

Supported Providers

• Downstream clients should call /v1/responses for generation. /v1/chat/completions, /v1/messages, and Google GenAI generation endpoints are not exposed as client-facing generation routes.
• DeepSeek, Opencode Go, and other services that expose an OpenAI Chat Completions-compatible upstream API. Rosetta performs protocol conversion and tool-layer translation.
• OpenAI, API relay services, and other services that expose an OpenAI Responses-compatible upstream API. Rosetta directly passes through these requests without decoding and re-encoding them.
• Anthropic Messages and Google GenAI upstream providers remain available through the conversion pipeline.

Citation

Contributing

Contributions are welcome. Visit the GitHub repository to get started.

License

This project keeps the MIT license. See LICENSE for details.