Open Interpreter

Type: full-code · Vendor: Open Interpreter · Language: Python · License: AGPL-3.0 · Status: active · Status in practice: mature · First released: 2023-07-14

Links: homepage docs repo

Open-source local code-execution agent: equips a function-calling LLM with an exec() function so it can run Python, JavaScript, Shell, and more on the user's machine, with default human approval and an experimental OS mode for visual desktop control.

Description. Open Interpreter runs as a terminal chat (`interpreter`) or as a Python library. Under the hood it gives a function-calling model an exec() function and lets it execute generated code locally across Python, JavaScript, Shell, and more. By default each block of code requires user approval before it runs; `auto_run` / `interpreter -y` bypasses that. OS mode (--os) is an experimental computer-use mode where the model captures screenshots and drives the mouse/keyboard. Safe mode is an experimental code-scanning layer with off/ask/auto settings.

Agent loop shape. ReAct-style local loop. Function-calling LLM emits a code block in a chosen language; the runtime asks the user to approve, then executes via exec() and feeds stdout/stderr/return values back. With auto_run on, approval is skipped. OS mode swaps the textual exec() loop for a screenshot/mouse/keyboard loop driven by a multimodal model.

Primary use cases

natural-language interface to the user's computer via local code execution
ChatGPT Code Interpreter-style workflows running locally
ad-hoc data, file, and OS automation across multiple languages
experimental visual desktop control via OS mode

flowchart TD user[User chat] --> llm[Function-calling LLM] llm --> codeblock[Code block in language] codeblock --> approve{auto_run?} approve -->|no| ask[Ask user] approve -->|yes| exec ask -->|approved| exec[exec()] ask -->|denied| llm exec --> obs[stdout / stderr / value] obs --> llm llm --> done[Done] osmode[OS mode --os] -.replaces exec() loop.-> screen[Screenshots + mouse/keyboard]

Key concepts

exec() over multiple languages → code-execution (docs) — Local function-calling LLM with an exec() function for Python, JavaScript, Shell, etc.
Approval per code block → approval-queue (docs) — Default behaviour: user approves each block before it runs.
auto_run (docs) — Bypass approval (`interpreter -y` or `interpreter.auto_run = True`).
OS mode → computer-use (docs) — Experimental visual desktop control via screenshots, mouse, and keyboard.
Safe mode — Experimental code-scanning with off/ask/auto settings.