What is Open Interpreter?
Open Interpreter began in 2023 as a Python tool that let a language model run code on your computer. The project has since been rewritten: the main repository is now a Rust fork of OpenAI's Codex CLI, aimed at low-cost and open models such as Kimi, GLM, DeepSeek and Qwen. The original Python version lives on as a community fork, endolith/open-interpreter.
Its distinguishing feature is harness emulation. A /harness command switches between 10 reimplementations of other agents' prompting and tool loops, including claude-code, kimi-code, qwen-code, deepseek-tui and swe-agent, on the theory that each model works best inside the harness its maker tuned it for. It speaks the Codex exec protocol and the Agent Client Protocol, reads AGENTS.md and shared .agents/skills folders, and supports MCP, hooks and sandboxed command execution. The repository has about 68,900 stars, is Apache 2.0 licensed and released rust-v0.0.61 on 11 November 2026; Interpreter Workstation is its desktop client.
Is Open Interpreter worth it?
Model makers tune their models against their own agent harness, and a model dropped into someone else's loop often does worse than its benchmarks suggest. Open Interpreter's rewrite is built on that observation. It forks OpenAI's Codex CLI, keeps its sandbox and exec protocol, and adds 10 Rust reimplementations of other harnesses, so Kimi runs inside a copy of Kimi Code, Qwen inside Qwen Code, and so on. The project does not publish benchmark numbers showing how much this helps, which is the one figure a reader would want.
The security model is Codex's, which is a good thing to inherit. Commands run in native sandboxes, Seatbelt on macOS and Bubblewrap with seccomp on Linux and WSL, with three modes (read-only, workspace-write and danger-full-access) and three approval policies. In workspace-write, network access is off unless you turn it on, and the docs say the tool should stop rather than run unsandboxed when a policy cannot be enforced. A --yolo flag removes both sandbox and approvals; the docs reserve it for throwaway VMs.
Portability is a deliberate goal. It reads AGENTS.md and shared .agents/skills directories, works in ACP editors through interpreter acp, and accepts the Codex SDK with a one-line binary override. Configuration and session state stay local under ~/.openinterpreter. Installation, though, is a curl-to-shell script from the project's site.
The name is the confusing part. The repository dates from July 2023 and has about 68,900 stars and 10,792 commits, but most of that history is the Python program people starred, now handed to a community fork. The Rust line is young, with version numbers still at 0.0.61 on 11 November 2026, though it ships often and carries only 41 open issues. Against Codex itself the difference is the harness switch and the wider provider catalogue; against opencode, a tighter focus on making cheap models behave.
Where it earns its place
A solo developer paying for API tokens out of pocket
Get Claude Code-like results from a cheaper open model
Install, add a Kimi or DeepSeek key, and use /harness to pick the model maker's own harness. The tool is free; cost is the provider's token price.
A team that already built tooling on the Codex SDK
Try open models without rewriting the integration
Set codexPathOverride to interpreter in the SDK constructor; the project ships a provider-free compatibility check script to confirm the protocol matches.
A developer whose editor supports the Agent Client Protocol
Use one agent inside the editor with a choice of providers
Configure the client to launch interpreter acp, keep workspace-write with on-request approvals, and switch models from the TUI or client with /model.
Before you commit
- Setup
- curl -fsSL https://www.openinterpreter.com/install | sh on macOS and Linux, or a PowerShell one-liner on Windows; type i or interpreter to start, then add a provider key or point at a local model.
- Integrations
- OpenAI-compatible providers and Chat Completions, Kimi, DeepSeek, Z.AI GLM and others; Agent Client Protocol editors; Codex SDK and exec protocol; MCP; hooks; AGENTS.md and .agents/skills; agent-browser and trycua for QA testing.
- Data and privacy
- Config and session state are stored locally under ~/.openinterpreter. Code and prompts go to whichever provider you select. Commands run in an OS sandbox; workspace-write disables network access unless enabled. Telemetry is not described in the sandbox docs.
- Cost at scale
- No licence cost; spending is entirely the token bill of the provider you choose, or none with a local model.
- Limits
- Rust line at 0.0.61; no published benchmarks for harness emulation; --yolo disables all protection; 41 open issues; the Python-era features and docs now live in a separate community fork.
What’s good
- Switches among 10 emulated harnesses tuned to specific model families
- Inherits Codex's sandboxing, approvals and exec protocol
- ACP, MCP, AGENTS.md and shared skills folders supported
- Codex SDK users can swap it in with one line
What’s not
- No published evidence that harness emulation improves results
- Rust rewrite is still at version 0.0.x
- Star count and history mostly reflect the old Python project
- Install path is a curl-to-shell script
Best for
Developers who want a Codex-style terminal agent but run Kimi, GLM, DeepSeek or Qwen models to keep costs down.
Open Interpreter alternatives
Open Interpreter: common questions
- Is Open Interpreter worth it?
- Open Interpreter is now a well-built Codex fork with one good idea, running cheap models inside the harness they were tuned for, and the 68,900 stars it carries mostly belong to a different, older program. We rate it 3.9 out of 5.
- How much does Open Interpreter cost?
- Free and open source (Apache 2.0). You bring a provider key or a local model; there is no Open Interpreter subscription in the terminal tool.
- What are the downsides of Open Interpreter?
- No published evidence that harness emulation improves results; rust rewrite is still at version 0.0.x; star count and history mostly reflect the old Python project; install path is a curl-to-shell script.
- Who is Open Interpreter best for?
- Developers who want a Codex-style terminal agent but run Kimi, GLM, DeepSeek or Qwen models to keep costs down.
- How much does Open Interpreter cost for a team or in production?
- No licence cost; spending is entirely the token bill of the provider you choose, or none with a local model.
- Is Open Interpreter safe for company data?
- Config and session state are stored locally under ~/.openinterpreter. Code and prompts go to whichever provider you select. Commands run in an OS sandbox; workspace-write disables network access unless enabled. Telemetry is not described in the sandbox docs.
- What are the best Open Interpreter alternatives?
- The closest alternatives we have reviewed are OpenAI Codex CLI, OpenCode, Qwen Code, Interpreter Workstation and Aider.
Answers checked Nov 21, 2026.