Open Interpreter
A terminal coding agent for low-cost models, with switchable Rust harnesses plus ACP and Codex exec compatibility.
Per-dimension scores and reasoning
Evidence shows support for native sandboxing (macOS/Linux/Windows), permissions system, hooks, approvals, and shared AGENTS.md and skills directories, indicating mechanisms for least privilege and user confirmation. However, implementation details are not provided, and it is unverified whether sandboxing is enabled by default or permissions are minimal by default. Data flow transparency is partial: README mentions product config and session state stored under ~/.openinterpreter but does not detail data flows. Sensitive data handling is not explicitly addressed. Dependency security: package.json includes version pins and overrides, but no vulnerability scanning or security audit evidence. External effects: supports executing commands and network operations, but default restrictions are unclear. Rollback is not mentioned. Source attribution: project is based on OpenAI Codex, but code provenance and contributors are not clearly stated. Therefore, most criteria scored 1, rollback 0.
Self-consistency: README descriptions align with documentation links, but no code-level verification. Dependency availability: many dependencies, but no lockfile or integrity checks. Failure messages: no examples of error handling or user prompts. Therefore, self-consistency scored 2, others 1.
Audience and scenarios: README targets developers and end-users, with multiple use cases (ACP, Codex SDK, QA skill). Capability boundaries: clearly lists supported features and limitations (e.g., sandboxing, permissions). Trigger precision: provides commands like /harness and /model, but detailed trigger conditions are not specified. Environment fit: supports macOS, Linux, Windows, with install scripts. Therefore, most scored 2, trigger precision 1.
Information architecture: README is well-structured with installation, features, and documentation links. Install notes: provides curl and PowerShell commands. Naming stability: project name and commands are stable, but no version history. Examples and FAQ: multiple documentation links and examples. Known limitations: not explicitly listed. License: Apache-2.0, but copyright notice is OpenAI, which may cause confusion. Versioning/changelog: not provided. Maintenance responsibility: maintainers or contribution guidelines not clearly stated. Therefore, most scored 2, naming stability, known limitations, versioning/changelog, maintenance responsibility 1.
Output usability: provides CLI and TUI, supports multiple models and providers. Marginal value: as a fork of Codex, adds harness emulation and ACP support. Cost-benefit: optimized for low-cost models, potentially reducing usage costs. Therefore, all scored 2.
Claim traceability: claims in README lack specific evidence or test results. Cross-source corroboration: no third-party verification. Fact-inference separation: does not clearly distinguish facts from inferences. Therefore, all scored 1.
- It is unverified whether sandboxing is enabled by default or permissions are minimal by default.
- Dependency security lacks evidence of vulnerability scanning or audits.
- Rollback mechanism is not mentioned, which may affect failure recovery.
- Copyright notice is OpenAI, but project is a fork; license compliance should be confirmed.
What does this agent do, and when should you use it?
Open Interpreter is a Rust-based terminal coding agent positioned to get stronger agent behavior from low-cost models. It is a fork of OpenAI Codex and lets users switch the active harness with /harness, including native, kimi-code, qwen-code, deepseek-tui, and claude-code. The agent runs commands in native sandboxing on macOS, Linux, and Windows, and supports exec, MCP, skills, hooks, permissions, and AGENTS.md. It can serve ACP-compatible editors and clients through interpreter acp, or replace the executable behind an existing Codex SDK integration through codexPathOverride. Product-specific configuration and runtime state live under ~/.openinterpreter, while repository instructions and skills can use AGENTS.md and .agents/skills.
After installation, i or interpreter starts a terminal session. In that session, /model changes the provider and model, while /harness selects a native or emulated harness; the listed harnesses include native, kimi-code, qwen-code, deepseek-tui, claude-code, and minimal. It executes commands in native sandboxing, reads repository AGENTS.md and shared .agents/skills directories, supports MCP, skills, hooks, and permissions, and stores product-only configuration and session state under ~/.openinterpreter. For ACP clients, it launches as interpreter acp. For an application already using the OpenAI Codex SDK, new Codex({ codexPathOverride: "interpreter" }) routes execution to Open Interpreter over the same Codex exec protocol. Its QA skill can test web applications with agent-browser or operate and test native applications with trycua.
- A developer using lower-cost hosted models for terminal coding can switch models and harnesses from one session.
- A team with an existing OpenAI Codex SDK integration can keep its SDK calls and set codexPathOverride to "interpreter".
- A developer using an ACP-compatible editor or client can configure it to launch interpreter acp.
- An engineering team testing a web or desktop application can use the included QA skill with agent-browser or trycua.
- A team standardizing repository instructions and reusable skills across compatible coding tools can retain AGENTS.md and .agents/skills.
What are this agent's strengths and limitations?
- Includes a Rust implementation of the provider-recommended Kimi Code harness behind a Codex-like interface.
- Lets one terminal agent switch among several named harness styles instead of fixing the workflow to one interface.
- Explicitly supports ACP and the Codex exec protocol; an existing Codex SDK integration has a documented one-line executable override.
- Runs commands in native sandboxing across macOS, Linux, and Windows, with documented web and native-app QA paths.
- Prefers shared conventions such as AGENTS.md, .agents/skills, and MCP instead of requiring new skills in a product-only directory.
- Live hosted-model sources require provider credentials, but the supplied material does not provide exact authentication configuration.
- The documented installation paths download and execute a remote shell or PowerShell script, which may conflict with restrictive installation policies.
- Web and native-app QA respectively use agent-browser and trycua, so those capabilities depend on the relevant tool path.
- This is a new Rust version distinct from the original Python project, which continues as the community-maintained endolith/open-interpreter fork; existing users need to assess that boundary.
How do you install or deploy this agent?
macOS or Linux: curl -fsSL https://www.openinterpreter.com/install | sh
Windows PowerShell: irm https://www.openinterpreter.com/install.ps1 | iexThen run i or interpreter to start a session. The supplied material says live model sources for hosted providers require the credentials documented in the provider guides, but it does not specify the exact credential or configuration values.
How do you use this agent?
Run interpreter in a terminal, then use /model to choose a provider and model and /harness to inspect or switch harnesses. To use it as an ACP agent, launch interpreter acp and configure the ACP-compatible client to start that command. In a Codex SDK application, replace new Codex() with new Codex({ codexPathOverride: "interpreter" }). Run scripts/test-codex-sdk-compat.sh for a local compatibility check that does not require a provider.
How does this agent compare with similar options?
Compared with OpenAI Codex, this repository identifies itself as a Codex fork focused on emulating high-performing agent harnesses for low-cost models while retaining Codex exec protocol compatibility. The original Python Open Interpreter project continues as the community-maintained endolith/open-interpreter fork; this repository is the newer Rust version.