free-code — A stripped-down build of Claude Code

free-code is a fork of Anthropic's Claude Code, a terminal program that acts as an AI coding assistant. The original source code became publicly available in March 2026 after being exposed through Anthropic's npm distribution. This project takes that source and modifies it in three ways.

The three changes

1. Telemetry removed — The original program sends usage data, crash reports, and analytics back to Anthropic's servers. This build removes or disables all of that. Nothing phones home.

2. Security-prompt guardrails removed — Anthropic injects extra system-level instructions into every conversation that restrict what the AI can do beyond its own built-in safety training. This build strips those injected restrictions. The model's own safety training still applies — only the extra layer is removed.

3. Experimental features unlocked — The original ships with 88 feature flags, most disabled. This build enables all 54 that compile successfully. These include voice input, deep-thinking mode, background automation triggers, memory extraction, and more.

Model providers

free-code works with five different AI providers, switched via environment variables:

Provider	How to enable	Auth method
Anthropic (default)	—	API key or OAuth
OpenAI Codex	CLAUDE_CODE_USE_OPENAI=1	OAuth via OpenAI
AWS Bedrock	CLAUDE_CODE_USE_BEDROCK=1	AWS credentials
Google Vertex AI	CLAUDE_CODE_USE_VERTEX=1	Google Cloud login
Anthropic Foundry	CLAUDE_CODE_USE_FOUNDRY=1	API key

How to install and use

Quick install (one command):

curl -fsSL https://raw.githubusercontent.com/paoloanzn/free-code/main/install.sh | bash

This checks your system, installs Bun (a JavaScript runtime) if needed, builds the project, and puts a free-code command on your PATH.

Requirements: Bun ≥ 1.3.11, macOS or Linux (Windows via WSL), and credentials for a model provider.

Basic usage:

free-code                    # Interactive chat mode
free-code -p "question"      # One-shot mode: ask a single question
free-code --model claude-opus-4-6   # Pick a specific model
free-code /login             # Log in to your provider

Building from source

Clone the repo, then:

bun build          # Production build
bun run build:dev:full   # Build with all 54 experimental flags

You can also enable individual flags, e.g. bun run ./scripts/build.ts --feature=ULTRAPLAN --feature=ULTRATHINK.

Tech stack

• Runtime: Bun
• Language: TypeScript
• Terminal UI: React + Ink (a React renderer for terminals)
• CLI parsing: Commander.js
• Schema validation: Zod v4
• Protocols: MCP (Model Context Protocol), LSP (Language Server Protocol)
• APIs: Anthropic, OpenAI Codex, AWS Bedrock, Google Vertex AI

Notable details

• IPFS mirror: A permanent copy of the repo is pinned on IPFS (a decentralized storage network) with a content ID, so the code survives even if the GitHub repo is taken down.
• Contributing: The project welcomes help restoring the 34 broken feature flags — many are close to compiling and just need small fixes.
• License note: The original Claude Code source belongs to Anthropic; this fork exists because the source was publicly exposed. The README says to use it at your own discretion.

Key caveats

• The README explicitly states that removing guardrails does not disable the model's own safety training — it only removes the extra prompt-level restrictions the CLI adds.
• The project is a fork of proprietary code that was publicly exposed, not officially released by Anthropic.
• The README lists specific model names (e.g., Claude Opus 4.6, GPT-5.3 Codex) as supported options.

Project structure (high level)

• src/entrypoints/cli.tsx — entry point
• src/commands.ts — slash commands
• src/tools.ts — agent tools (Bash, Read, Edit, etc.)
• src/QueryEngine.ts — the LLM query engine
• src/screens/REPL.tsx — main interactive UI
• src/services/ — API clients, OAuth, analytics
• src/voice/ — voice input
• src/bridge/ — IDE integration (VS Code, JetBrains)

Bottom line

free-code is a community fork of Claude Code that removes data collection, removes extra safety restrictions, and unlocks hidden experimental features — while keeping the core functionality of a terminal AI coding assistant and adding support for multiple AI providers beyond Anthropic's own API.