Run coding tasks with low-cost models in a terminal agent
Typing i in the terminal opens a session where models like Kimi K3 or GLM 5.3 write code and execute commands. This Rust-based tool is a reimplementation of OpenAI's Codex, designed to maximize the performance of low-cost models. It integrates naturally into existing development environments, and its adherence to standard formats makes it easy to migrate to other agents.
Kimi K3 support and Rust reimplementation
A recent update reimplemented the provider-recommended Kimi Code harness in Rust to support Kimi K3. This allows you to leverage K3's performance through an interface similar to Codex. This version is a new Rust build based on Codex, not the original Python project. The original Python project remains as a separate fork maintained by the community.
Harness switching and protocol compatibility
You can instantly switch between various agent harnesses such as native, claude-code, and kimi-code by entering the /harness command in the terminal.
- Harness list It offers various options including native, claude-code, zcode, kimi-code, qwen-code, deepseek-tui, swe-agent, and minimal.
- SDK integration Specifying
interpreterin thecodexPathOverrideoption of the Codex SDK allows you to switch while maintaining your existing code structure.
Portability and standard format compliance
The tool is designed to avoid being locked into specific formats and to integrate naturally with existing agent configurations. It prioritizes support for standardized formats such as the repository's AGENTS.md, shared .agents/skills directory, MCP, ACP, and Codex exec protocols. User-written data is kept in readable files, simplifying the process of migrating to or importing from other compatible agents. Only product-specific settings and runtime state are stored in ~/.openinterpreter, while previous version skill directories remain readable for compatibility.
Computer use and sandboxed execution
Built-in QA skills allow the model to run web apps in a real browser or test native apps. It handles the entire process of manipulating and verifying interfaces using agent-browser or trycua. All commands are executed within native sandboxes on macOS, Linux, and Windows to ensure safety. You can freely switch providers and models using the /model command in the TUI, and OpenAI-compatible providers are supported via the --chat-completions flag.
Pre-adoption considerations
Since this tool focuses on maximizing the performance of low-cost models, it is important to select a harness tailored to the characteristics of those models. Being Rust-based, it may differ from workflows that rely on the existing Python ecosystem. Installation uses a curl script for macOS and Linux, and a PowerShell script for Windows. It is distributed under the Apache-2.0 license, and using live model sources requires credentials specified in the provider documentation. The provider list is not maintained as a Rust list but is generated by python3 scripts/write_provider_catalog.py, so you should verify if the script has been run to ensure support for the latest models.
