Open Interpreter rewrites itself in Rust to cut agent costs

Open Interpreter rewrote itself in Rust and repositioned as a coding agent optimized for low-cost models. The project, a fork of OpenAI’s Codex, now ships a harness emulation system that lets users switch between the agent protocols of Claude Code, Kimi CLI, Qwen Code, DeepSeek TUI, and others from a single terminal interface.

The move is a bet that the economics of AI coding agents have shifted. When Open Interpreter launched as a Python project in 2023, the assumption was that users would run it against expensive frontier models like GPT-4. The new Rust version flips that premise. It is designed to extract the best performance from models that cost pennies per million tokens. The harness system is the key mechanism.

Type /harness in the new terminal UI, and Open Interpreter lists eight active harnesses: native, claude-code, claude-code-bare, zcode, kimi-cli, qwen-code, deepseek-tui, swe-agent, and minimal. Each harness emulates the agent loop, tool-calling format, and system prompt of a different coding agent. The idea is that a model like Qwen-2.5-Coder-32B, which is tuned for Qwen Code’s agent format, will perform better when the harness matches its training distribution. Open Interpreter lets the user pick the harness that fits the model, not the other way around.

This is a practical insight that most agent frameworks ignore. Model providers ship their own agent implementations because the model’s instruction-tuning is baked into a specific conversation format. Swap the format and the model degrades. Open Interpreter’s harness abstraction acknowledges that the agent loop is not a generic scaffold. It is a runtime that must match the model’s training.

The Rust rewrite is not cosmetic. The original Python version was slow at startup and heavy at runtime, especially when running local models. The new version uses native sandboxing on macOS, Linux, and Windows, and runs commands inside a Rust-based execution environment. The project claims the rewrite makes the agent feel “instant” compared to the Python version, though the README does not publish benchmarks. The performance gain matters most for low-cost models, where every millisecond of overhead cuts into the cost advantage.

Open Interpreter also ships a QA skill that lets any model drive web apps through a real browser via agent-browser, or test native apps through trycua. This is a computer-use capability that does not depend on a specific model’s vision or action-tuning. Any model can use it, as long as the harness provides the right tool schema.

The project is Apache-2.0 licensed and keeps all config and session state local under ~/.openinterpreter. It supports exec, MCP, skills, hooks, permissions, and AGENTS.md files. It also runs as an Agent Client Protocol agent for editors via interpreter acp.

The original Python project now lives as a community-maintained fork at endolith/open-interpreter. The Rust version is the new mainline.

What this means for AI builders is straightforward. The cost of running coding agents has dropped to the point where the bottleneck is no longer the model’s per-token price. The bottleneck is the agent harness. Open Interpreter’s harness system lets a developer run the same task against Claude Code’s protocol one minute and Qwen Code’s the next, without changing the underlying model. That is useful for testing, but it is also a hedge. If the model landscape shifts, the harness does not need to change.

The project’s focus on low-cost models also signals a broader market shift. Coding agents have been a premium product, sold as monthly subscriptions or priced by the token. Open Interpreter is free, local, and designed to run on consumer hardware. It does not require an API key, though it supports provider switching via /model. A developer on a laptop can run a coding agent against a local Qwen model and pay nothing in inference costs.

The tradeoff is capability. Low-cost models are not Claude Code or GPT-4. They make more mistakes, follow instructions less reliably, and produce lower-quality code. Open Interpreter does not solve that. It optimizes the harness for the models that exist, not the models that do not. The project’s README is honest about this. It does not claim parity with frontier agents. It claims efficiency.

The harness emulation approach also raises a question about the sustainability of model-specific agent formats. If every model provider ships a proprietary agent protocol, the ecosystem fragments. Open Interpreter’s answer is to emulate them all. That is a reasonable engineering response, but it creates a maintenance burden. Every time Anthropic updates Claude Code’s system prompt or Kimi changes its tool schema, Open Interpreter must update its harness.

The project’s longevity will depend on whether the community treats harnesses as a shared resource or a competitive moat. For now, Open Interpreter is the only open-source project that lets a developer run a Qwen model through a Claude Code harness and a DeepSeek model through a Kimi harness from the same terminal. That is a useful piece of infrastructure, and it costs nothing to try.