Open Interpreter Rewrote Itself in Rust to Chase Cheap Models
Open Interpreter is trending on GitHub again with 633 stars today, sitting on 66k total. Version 0.0.26 shipped July 16. But the project you remember is gone. This one is a Rust fork of OpenAI's Codex, and its entire reason for existing is one sentence in the README: a coding agent optimized for low-cost models.
The bet underneath is worth spelling out. Everyone assumes the model is what makes a coding agent good, so the agents get built around frontier models and everything cheap looks broken by comparison. Open Interpreter's claim is that a big chunk of that gap is the harness, not the model. Feed Kimi K3 or Qwen or DeepSeek into a loop designed for them and you close more of the distance than the benchmark tables suggest.
So the interesting feature is harness switching. You can run native, or claude-code, or kimi-cli, and pick whichever loop squeezes the most out of the model you're actually paying for. That's an unusual thing to ship. Most agents treat their harness as the product and the model as a swappable backend. This one inverts it: the model is the constraint you're given, the harness is the variable you tune.
The rest is what you'd expect from a serious terminal agent. Sandboxed command execution, computer use for driving web and native apps during testing, Agent Client Protocol support, Apache 2.0, install it and run `i`.
Timing is the tell. This landed the same day Kimi K3 went live, and K3's open weights drop July 27. A near-frontier open model is about to be free, and the harness that makes it usable on your own hardware just shipped. That's not luck, that's someone reading the calendar.
https://github.com/openinterpreter/openinterpreter
← Back to all articles
The bet underneath is worth spelling out. Everyone assumes the model is what makes a coding agent good, so the agents get built around frontier models and everything cheap looks broken by comparison. Open Interpreter's claim is that a big chunk of that gap is the harness, not the model. Feed Kimi K3 or Qwen or DeepSeek into a loop designed for them and you close more of the distance than the benchmark tables suggest.
So the interesting feature is harness switching. You can run native, or claude-code, or kimi-cli, and pick whichever loop squeezes the most out of the model you're actually paying for. That's an unusual thing to ship. Most agents treat their harness as the product and the model as a swappable backend. This one inverts it: the model is the constraint you're given, the harness is the variable you tune.
The rest is what you'd expect from a serious terminal agent. Sandboxed command execution, computer use for driving web and native apps during testing, Agent Client Protocol support, Apache 2.0, install it and run `i`.
Timing is the tell. This landed the same day Kimi K3 went live, and K3's open weights drop July 27. A near-frontier open model is about to be free, and the harness that makes it usable on your own hardware just shipped. That's not luck, that's someone reading the calendar.
https://github.com/openinterpreter/openinterpreter
Comments