Open Interpreter Review

The README describes Open Interpreter as a coding agent optimized for low-cost models, a Rust version based on OpenAI's Codex, and says it is a fork focused on harness emulation. Commands run inside native sandboxing on macOS, Linux, and Windows, and the tool supports exec, MCP, skills, hooks, permissions, and AGENTS.md. Harnesses are switched with /harness and emulate agent setups that perform well on cheaper models: native, claude-code, claude-code-bare, zcode, kimi-code, kimi-cli, qwen-code, deepseek-tui, swe-agent, and minimal. /model switches models in the TUI, the README names Kimi K3 and GLM 5.3 and has provider guides for Kimi K3 (using a Rust reimplementation of the provider-recommended Kimi Code harness), DeepSeek, and Z.AI or GLM, and any OpenAI-compatible provider works through Chat Completions with --chat-completions. Because it speaks the Codex exec protocol, the Codex SDK can switch to it with a one-line codexPathOverride change, and it adds ACP support via interpreter acp. A built-in QA skill can drive web apps in a browser through agent-browser and test native apps through trycua.

Install with the curl script from openinterpreter.com on macOS and Linux or the PowerShell script on Windows, then start a session by typing i or interpreter. Repository instructions live in AGENTS.md, new skills belong in .agents/skills or ~/.agents/skills (legacy product-specific skill directories remain readable), and product-only configuration and session state live in ~/.openinterpreter. The README links to a Sandbox and approvals doc but does not describe the approval model, so read it before running unattended, and it links FORK_BRANDING.md for distribution forks. Provider and model lists are generated by a script from the codex-rs code and live sources need the provider credentials described in the provider docs. The original Python implementation is maintained as a community fork at endolith/open-interpreter. OpenAlternative lists the version as v0.0.56, and the project's own site page we fetched contained only its title, Coding agent for open models.

It behaves like other terminal coding agents: you type a task, the agent proposes and runs commands in a sandbox, and you review. Output quality depends heavily on the model you pick, and the project aims to work well with cheaper and open-weight ones, using harness emulation to match how those models were trained to work. The README has no limitations section. This section reflects the README and OpenAlternative; we did not run it.

• Rust coding agent forked from the Codex codebase, speaking the Codex exec protocol
• Native sandboxing on macOS, Linux, and Windows
• MCP, skills, hooks, permissions, AGENTS.md, and ACP support
• Switchable harnesses such as native, claude-code, kimi-code, qwen-code, deepseek-tui, and swe-agent
• Any OpenAI-compatible provider, with guides for Kimi K3, DeepSeek, and Z.AI GLM
• Built-in QA skill for driving web and native apps

• Targets low-cost and open-weight models, which keeps usage costs down
• Sandboxed execution across all three desktop operating systems
• Apache-2.0 license and a large inherited community, with about 68,500 stars
• Drop-in compatibility with the Codex SDK and exec protocol

• The project's identity changed: the repository is no longer the original Python Open Interpreter, so older tutorials may not apply
• A 0.0.x version number and a recent rewrite imply rapid change
• Quality depends on the model, and small low-cost models can struggle on hard tasks
• The README does not describe the approval model and lists no limitations
• Fewer enterprise controls than vendor-supported coding agents

Open Interpreter is free and open source under Apache-2.0. Costs come from the model providers you connect, and the project positions itself around low-cost models and not requiring a proprietary subscription.

Claude Code, Codex, and Devin are vendor-built coding agents tied to their own models and pricing. Open Interpreter is an open agent, forked from the Codex codebase, that is tuned for cheaper and open-weight models and can switch harnesses. Compared with OpenCode, Pi, and Cline, it leans harder into low-cost model support and native sandboxing, with a younger track record in its new form.

• Developers who want to run a coding agent on cheap or open-weight models
• Users who want native sandboxing out of the box
• People who want to switch between harnesses and models quickly

• Teams that want the original Python Open Interpreter, which now lives in a community fork
• Developers who need the strongest available models and tight vendor integration
• Anyone who needs a documented, stable approval model today