claude-clean 0.1.0

CLI for selectively purging Claude Code project data — history, settings, metadata, or everything at once.

Claude Clean

Selectively purge Claude Code project data from your local machine.

Claude Code stores conversation history, session artifacts, project settings, and memory files locally under ~/.claude/ and ~/.claude.json. Over time, this data accumulates across every project you open. claude-clean gives you precise, surgical control over what to keep and what to remove — per project, per data category, or everything at once.

[!CAUTION] All deletions performed by this tool are permanent and irreversible. There is no undo, no trash, and no recovery mechanism. Always run with --dry-run first to review the exact list of files and directories that will be affected before committing to a destructive operation.

---

Table of Contents

• Quick Start
• Installation
• Usage
  • Project Selection
  • Commands
  • Global Options
• What Gets Deleted
• Supported Platforms
• Development
• License

---

Quick Start

The fastest way to run Claude Clean without installing it globally:

uvx claude-clean purge --dry-run

This launches an interactive project selector, shows you everything that would be deleted, and exits without touching the filesystem.

---

Installation

Requires Python 3.14 or later.

Recommended: Run without installing (via uvx)

uvx (part of uv) runs Python CLI tools in isolated, ephemeral environments — no global install needed:

# Run any command directly
uvx claude-clean history --all --dry-run

# Equivalent to installing + running, but leaves no footprint
uvx claude-clean purge --project /path/to/project --yes

This is the recommended approach for one-off or occasional use.

Traditional: Install from PyPI

pip install claude-clean

After installation, the claude-clean command is available globally:

claude-clean --help

From source

git clone https://github.com/j-about/Claude-Clean.git
cd Claude-Clean
pip install .

---

Usage

Project Selection

Every command needs to know which project(s) to operate on. Claude Clean discovers projects by reading the "projects" key from ~/.claude.json — the same state file Claude Code uses internally.

There are three ways to select projects:

Method	Flag	Behavior
Interactive	(default)	Displays a numbered list of all known projects with an "ALL" option. You pick one or all.
Explicit	--project PATH / -p PATH	Targets a single project by its absolute path. The path must match a key in ~/.claude.json.
Blanket	--all / -a	Targets every known project without prompting.

Interactive mode example:

$ claude-clean history

  Known Projects
  ┌───┬─────────────────────────────────────┐
  │ # │ Project Path                        │
  ├───┼─────────────────────────────────────┤
  │ 1 │ /home/user/project-alpha            │
  │ 2 │ /home/user/project-beta             │
  │ 3 │ ALL                                 │
  └───┴─────────────────────────────────────┘

  Select a project (1-3):

Commands

claude-clean history

Delete conversation history and all associated session artifacts for selected project(s).

What it removes:

• Matching entries from ~/.claude/history.jsonl
• Cached paste content referenced by removed entries (~/.claude/paste-cache/*.txt)
• Session file-history backups (~/.claude/file-history/{session-id}/)
• Session environment snapshots (~/.claude/session-env/{session-id}/)
• Plan documents referenced in session files (~/.claude/plans/*.md)
• The encoded project directory (~/.claude/projects/{encoded-path}/)

# Interactive project selection, preview only
claude-clean history --dry-run

# Delete history for a specific project
claude-clean history --project /home/user/my-project --yes

# Delete history for all projects, skip confirmation
claude-clean history --all --yes

---

claude-clean settings

Delete project settings at the project scope, user scope, or both.

Scopes:

Scope	What it removes
project	The {project-root}/.claude/ directory (project-local settings, rules, commands, skills, agents, MCP config)
user	The project's entry from the "projects" object in ~/.claude.json (user-level state: allowed tools, MCP servers, session metrics)
all	Both of the above

When --scope is omitted, an interactive prompt asks you to choose.

# Interactive project + scope selection
claude-clean settings

# Delete project-local settings only
claude-clean settings --project /home/user/my-project --scope project --yes

# Remove user-level state for all projects
claude-clean settings --all --scope user --yes

# Full settings cleanup, preview first
claude-clean settings --all --scope all --dry-run

---

claude-clean metadata

Delete project memory files (CLAUDE.md).

What it removes:

• {project-root}/CLAUDE.md
• {project-root}/.claude/CLAUDE.md

# Delete memory files for a specific project
claude-clean metadata --project /home/user/my-project --yes

# Preview metadata cleanup for all projects
claude-clean metadata --all --dry-run

---

claude-clean purge

Run all of the above in sequence — history, then settings (scope: all), then metadata — as a single operation. This is the nuclear option for completely removing every trace of a project from Claude Code's local data.

# Full purge for a single project
claude-clean purge --project /home/user/my-project --yes

# Preview a full purge of all projects
claude-clean purge --all --dry-run

# Full purge of everything, no prompts
claude-clean purge --all --yes

Global Options

These options are available on every command:

Option	Short	Description
--project PATH	-p	Target a specific project by its absolute path
--all	-a	Target all known projects
--yes	-y	Skip the confirmation prompt
--dry-run		Preview the list of actions without performing any deletions
--help		Show help for the command

Recommended workflow:

# Step 1: Always preview first
claude-clean purge --project /path/to/project --dry-run

# Step 2: Review the output, then execute
claude-clean purge --project /path/to/project --yes

---

What Gets Deleted

The following table maps each command to the Claude Code data it targets:

Data	Location	history	settings --scope project	settings --scope user	metadata	purge
History entries	~/.claude/history.jsonl	X				X
Paste cache	~/.claude/paste-cache/{hash}.txt	X				X
Session transcripts	~/.claude/projects/{encoded}/	X				X
File-history backups	~/.claude/file-history/{session}/	X				X
Session environment	~/.claude/session-env/{session}/	X				X
Plan documents	~/.claude/plans/{name}.md	X				X
Project-local settings	{project}/.claude/		X			X
User-level project state	~/.claude.json (projects key)			X		X
Root memory file	{project}/CLAUDE.md				X	X
Nested memory file	{project}/.claude/CLAUDE.md				X	X

---

Supported Platforms

Platform	Home directory	Status
macOS 13.0 Ventura+	/Users/{username}	Supported
Linux Ubuntu 20.04+, Debian 10+, Fedora, etc.	/home/{username}	Supported
Windows 10 (build 1809+) / 11 (native)	C:\Users\{username}	Supported
Windows WSL2	/home/{username} (inside WSL filesystem)	Supported

claude-clean uses pathlib.Path throughout and resolves the home directory via Path.home(). No platform-specific APIs or hardcoded paths are used.

---

Development

Prerequisites

• uv (recommended) or pip
• Python 3.14+

Setup

git clone https://github.com/j-about/Claude-Clean.git
cd Claude-Clean
uv sync --group dev

Running Tests

uv run pytest tests/ -v

Type Checking

uv run mypy --strict src/

Linting and Formatting

uv run ruff check src/ tests/
uv run ruff format src/ tests/

Project Structure

src/claude_clean/
    __init__.py       Entry point — re-exports main()
    cli.py            Typer application, commands, action execution engine
    services.py       Pure functions returning planned actions (no side effects)
    utils.py          Path encoding, project discovery, interactive selection
    models.py         Data models — Action, ActionKind, ClaudeDataPaths
tests/
    conftest.py       Shared fixtures (realistic ~/.claude/ tree under tmp_path)
    test_utils.py     Path encoding, project discovery, interactive selection
    test_services.py  Action planning for all commands
    test_cli.py       CLI integration tests via CliRunner

Design Principles

• Separation of I/O and logic. Service functions accept paths and return lists of Action objects describing filesystem mutations. The CLI layer handles display, confirmation, and execution. This makes the core logic trivially testable with no mocks.

• Atomic writes. All JSON and JSONL rewrites use tempfile.mkstemp() in the same directory followed by os.replace(), ensuring no data corruption on crash or interrupt.

• Batch efficiency. When operating on multiple projects, shared files like history.jsonl and ~/.claude.json are rewritten exactly once — not once per project.

---

License

MIT — Copyright (c) 2026 Jonathan About