Authority AI Editorial
Open Interpreter: running code on your own machine
Open Interpreter is a coding agent that lives in your terminal, reading a project's files, proposing edits, and running commands on the machine it is installed on. This is worth stating carefully because the project has changed since it was first known as a Python tool for turning natural language into local code execution: the current version is a Rust rewrite forked from OpenAI's Codex agent, built for developers who want a terminal-based coding assistant tuned to work well with lower-cost and open models, not only frontier ones.
What it does
The agent reads the files in a project directory, plans a change from a plain-language request, edits code, and executes commands, keeping the developer in the loop for approval rather than acting unsupervised. It is provider-agnostic: a session can be pointed at OpenAI, Anthropic, local models, or other providers, and switched between them mid-session with a /model command. A separate /harness command switches the agent harness itself, the strategy the agent uses to plan and act, with named harnesses for models such as Kimi Code and DeepSeek alongside a Claude Code-style harness, which is the project's way of getting strong results out of models that are not OpenAI's own. It also runs as an Agent Client Protocol agent, so it can be driven from inside a compatible editor rather than only a standalone terminal, and it ships a QA skill for testing web and native applications.
The harness idea is the part worth sitting with, because it is the reason the rewrite exists at all. A single fixed prompting strategy tends to get uneven results across models with very different strengths, so instead of one house style for every model, Open Interpreter ships several, each tuned to how a particular model or model family reasons and calls tools best. The README specifically calls out a reimplemented, provider-recommended harness for Kimi K3, described as giving that model its strongest performance with a Codex-like interface on top. The project is also built to be compatible with OpenAI's Codex SDK through a single-line override, which lets a team keep existing Codex-based tooling while swapping in Open Interpreter's harness layer underneath it.
How you use it
Installation is a single shell command, curl -fsSL https://www.openinterpreter.com/install | sh on macOS or Linux, or a PowerShell equivalent on Windows, after which the interpreter command (or the short alias i) is available. A typical session starts from inside a project directory: run interpreter, describe the change in plain language, and the agent proposes and, on approval, makes the edit or runs the command. Safety is layered rather than a single switch: a sandbox determines what the agent is technically able to execute, and a separate approval mode determines when it must stop and ask before acting, so a developer can tighten or loosen either independently of the other.
What to know before adopting it
The project is Apache-2.0 licensed. The README is explicit about the rewrite: it describes the current codebase as "the new Rust version of Open Interpreter, based on Codex," and points anyone who wants the original Python project to a community-maintained fork at endolith/open-interpreter, which is a separate, independently maintained project going forward. There is no published pricing for the tool itself; cost depends entirely on whichever model provider a session is configured to use. Because the harness-switching design is central to how it gets good results from cheaper models, adopting it well means picking a harness deliberately rather than accepting whatever loads by default.
Anyone who last looked at Open Interpreter years ago and remembers it purely as a Python natural-language shell should treat that memory as out of date. The maturity model has also shifted along with the codebase: this is now a Rust binary with an editor-integration protocol and a QA skill built in, closer in spirit to a production coding-agent CLI than to the original single-purpose script. Teams that standardized on the old Python behavior should evaluate the current release on its own terms, or move to the community fork if the original design is what they specifically need.
Where it sits next to similar tools
Open Interpreter now sits closest to other terminal-native, editor-integrated coding agents such as Codex itself, rather than to the general "run any Python from a chat prompt" tools it once resembled. Its distinguishing feature is the harness abstraction: instead of one fixed way of prompting and acting, it swaps in different, provider-recommended strategies depending on the model behind it. Anyone assembling the instruction layer that sits on top of a coding agent like this, rather than the harness itself, can draw on the AI Ultimate Skill Collection, which packages 3,029 ready-made skills in 12 packs that your agent follows step by step.