Open Interpreter.

Consider which version you are installing

Open Interpreter has changed since the early Python project. The current official version is a terminal agent on Rust, created on the basis of Codex and designed to choose different providers. The old Python code lives separately in the community. Therefore, the old installation manual or demonstration can describe a different product process.

This card covers the current interpreter command and its documentation. The project’s independent role is configuring model access and the agent environment, rather than providing its own language model.

Work within the project

The agent runs in the project folder, reads files, suggests changes, and executes commands. The environment uses the instructions AGENTS.md, skills and MCP. ACP is provided for compatible editors. Sessions can be continued, and the model and method of its connection are selected in the settings.

The input query is better associated with the observed result. Instead of the general “improving the project” specify a specific error, the limit of permissible changes and the method of verification. This helps to maintain control even with a long task in which the agent has to consistently look for context and change multiple files.

Replay the error first

For example, a function mishandles a date at a month boundary. Ask the agent to find the cause, show a minimal reproducible input and prepare a focused fix without changing the public interface. Request a brief list of changed files and checks performed.

After the agent finishes, open the diff. Make sure the fix addresses the reported case without changing neighboring rules. A command exiting successfully confirms only that particular run; other behavior may remain untested. This scenario illustrates how to frame a task and does not claim hands-on testing of the agent.

Authority and model

The documentation distinguishes the sandbox from the approval mode. One limits technically available actions; the other determines when the agent must ask for permission. Read these settings before exposing work secrets or commands with external effects.

You can use a remote provider or local model. Local file work does not in itself mean local calculation of the response: when a cloud connection context is passed to the provider. The behavior of the tools also depends on the capabilities of the chosen model.

Self-start budget

The open source code is distributed on Apache-2.0, but the calculations remain a separate expense item. Local model weights need memory and suitable hardware; the API provider considers use by its own rules. For a first acquaintance, choose a small project, an understandable amount of work and limited permissions. This makes it possible to assess the result and costs without a complex change in the entire working environment at once.

Suitable tasks

• Make a limited change to the codebase
• Explore a project with the selected model
• Automate the repetitive stage of work through CLI
• Connect an agent to a compatible editor