Purpose
Open Interpreter is a coding agent for open-source, low-cost models. It is a fork of OpenAI Codex, and its core goal is to emulate an agent harness that maximizes the performance of low-cost models. It is officially positioned as "A coding agent for open models like Kimi K3 and GLM 5.3".
Main Capabilities
- Optimized for low-cost models: Through harness emulation, models such as Kimi K3 and GLM 5.3 achieve better performance. It is officially stated that the provider-recommended Kimi Code harness has been reimplemented in Rust, providing a Codex-like interface.
- Harness switching: Use the /harness command to switch the currently active harness, and to inspect or switch the Rust native model harness.
- Model and provider switching: Use /model in the TUI to switch providers and models; you can also run any selected OpenAI-compatible provider (via Chat Completions) with interpreter --chat-completions or interpreter exec --chat-completions.
- Native sandbox execution: Run commands in a native sandbox on macOS, Linux, and Windows.
- Editor integration: Run as an Agent Client Protocol (ACP) agent, usable in ACP-compatible editors and clients, with the client configured to launch interpreter acp.
- Codex compatibility: Compatible with the OpenAI Codex SDK, allowing you to keep the SDK and do a one-line binary override; Open Interpreter uses the same Codex exec protocol.
- Built-in QA skill: Comes with a QA skill that lets any model operate and test interfaces, able to use agent-browser to drive web applications in a real browser, or use trycua to operate and test native applications.
- Portability: Prioritizes shared, tool-neutral standardized directories and formats, including repository AGENTS.md, shared .agents/skills directory, MCP, ACP, and the Codex exec protocol; product-specific storage is located at ~/.openinterpreter, used only for configuration and runtime state that does not yet have a practical shared standard. Legacy product-specific skill directories can still be read for compatibility, but new skills should be placed in .agents/skills or ~/.agents/skills.
- Other support: Supports exec, MCP, skills, hooks, permissions, and AGENTS.md.
Usage
After installing on macOS and Linux, enter i or interpreter in the terminal to start a session. You can switch providers and models with /model, and switch harnesses with /harness. To use it in an ACP-compatible editor, configure the client to launch interpreter acp. If you already use the OpenAI Codex SDK, you can do a one-line binary override to reuse the SDK. Official resources also provide a local streaming web chat guide, as well as Interpreter Workstation for full desktop, browser, or headless clients.
Use Cases
Suitable for developers who want to run a coding agent in a local sandbox, use low-cost open-source models (such as Kimi K3 and GLM 5.3) for coding tasks, integrate an agent within an ACP-compatible editor, or reuse the Codex SDK and exec protocol. It is also suitable for scenarios that require letting models operate and test web or native application interfaces.
Limitations
Official materials do not provide specific pricing, cost, platform support details, performance data, or effectiveness guarantees. Provider and model membership is generated rather than maintained as a Rust list; refreshing all managed providers from codex-rs requires running python3 scripts/write_provider_catalog.py, or repeating --provider <provider-id> to update only selected provider entries; live model sources require provider credentials documented in the provider documentation. The original Python project now exists as a community-maintained fork at endolith/open-interpreter.
