claude-interpreter

Open Interpreter-like REPL powered by Claude Code.

Send prompts, Claude writes Python, we execute it. No tool calls, no overhead — just code.

Install

# Requires claude CLI
npm install -g @anthropic-ai/claude-code

# Install claude-interpreter
pip install -e .

Usage

# Interactive REPL
claude-interpreter

# Short alias
ci

# Single prompt
ci "list all python files in this directory"

# Use Opus
ci -m opus "refactor this code"

# Continue last session for this directory
ci -c

# Safe mode (confirm before running code)
ci --safe

# Verbose (show cost + timing)
ci -v "analyze data.csv"

How it works

1. You type a prompt
2. Sent to Claude via claude -p (print mode, no built-in tools)
3. Claude responds with Python code blocks
4. Code is executed locally with full system access
5. Output is fed back to Claude for the next turn

State persists across code blocks. Sessions are saved per-directory — use -c to resume.

REPL Commands

Command	Description
exit / q	Quit
!<cmd>	Run shell command
?	Help
clear	Clear screen
reset	Reset session
cost	Show total cost
"""	Multiline input

Architecture

claude_interpreter/
  core/
    interpreter.py   # REPL + claude -p integration + session persistence
    executor.py      # Python code execution with stdout capture
    ui.py            # Rich terminal UI

Only dependency: rich (for the terminal UI). Everything else is Claude Code.

Session Persistence

Sessions are saved per working directory (~/.claude_interpreter/). Use ci -c to continue where you left off. reset in the REPL clears the saved session.