openinterpreter/openinterpreter

What it solves

Open Interpreter provides a coding agent that allows LLMs to run code locally on a user's machine. It specifically focuses on optimizing performance for low-cost models by emulating various agent harnesses and maintaining compatibility with open standards, preventing users from being locked into a single proprietary format.

How it works

It acts as a fork of OpenAI's Codex, implementing a system that can execute commands within native sandboxes on macOS, Linux, and Windows. It supports multiple model providers via Chat Completions and allows users to switch between different model harnesses (such as Kimi Code, Claude Code, or DeepSeek) to maximize the efficiency of the chosen model. It also integrates with the Agent Client Protocol (ACP) and the Codex exec protocol to work across different editors and clients.

Who it’s for

It is designed for developers and users who want an AI coding assistant that can interact directly with their local system, as well as those building agentic workflows who require a portable, tool-neutral setup.

Highlights

• Multi-Harness Support: Ability to switch between various provider-recommended harnesses to optimize low-cost model performance.
• Native Sandboxing: Executes commands securely across major operating systems.
• Cross-Client Compatibility: Works with ACP-compatible editors and the Codex SDK.
• Computer Use: Includes a QA skill for operating and testing web apps (via agent-browser) and native apps (via trycua).
• Portable Design: Uses shared standards like AGENTS.md and .agents/skills to ensure data is not trapped in a proprietary format.