CodeXe

CodeXe is a community fork of OpenAI Codex with native local Xenolith inference. It preserves Codex's agent loop, tools, sandbox and approvals, while adding local inference status and durable KV sessions.

Quick start

On Linux x86_64:

curl -fsSL https://xenolith.it/install.sh | sh -s -- --interface codexe

The installer sets up CodeXe and Xenolith, guides you through model configuration or reuses your existing setup, and starts CodeXe.

For later sessions, run this from your project directory:

codexe

Build from source

On Linux/WSL, install Rust 1.95.0, Python 3.11+, a C/C++ compiler, pkg-config, and OpenSSL, libcap and ALSA development headers (libssl-dev, libcap-dev, libasound2-dev on Ubuntu). Build Xenolith and obtain a supported model, then:

git clone https://github.com/crrisci/codexe.git
cd codexe
python3 scripts/install_codexe.py
xenolith config init --model /absolute/path/to/model.gguf
codexe

Skip config init if Xenolith is already configured. The engine must be on PATH, or selected with XENOLITH_BIN=/absolute/path/to/xenolith. The source installer builds the complete package and installs codexe in ~/.local/bin. See build and install for custom destinations, prebuilt local packages and runtime dependencies, or the container build guide for standalone Linux archives.

Native inference

• Live status: engine startup, prefill progress, generation speed and cache reuse.
• Service discovery: model, context, output limit and reasoning capabilities.
• Vision: image attachments and view_image with a configured projector.
• Reasoning: streamed thoughts and history across tools and resumed sessions.
• Durable sessions: checkpoints, KV restoration and interrupted-turn recovery.
• Rewind: an offline controller previews the context rebuild cost before applying it.

Xenolith is the default provider on Linux/WSL. Explicit provider selections take precedence. CodeXe uses ~/.codexe for configuration and conversations; CODEX_HOME can select another home.

The native integration supports PNG/JPEG images, one per user message or tool result. See current limits for tool compatibility, context recovery and platform restrictions.

• Engine configuration, images and reasoning
• Sessions and inference status
• Offline rewind
• Contributing · Upstream maintenance

License and attribution

CodeXe is licensed under Apache-2.0, maintained independently of OpenAI and unaffiliated with Intel. NOTICE preserves OpenAI and third-party and CodeXe attributions; see the license overview. The upstream README documents the integrated OpenAI Codex revision.

Project website: xenolith.it.