Claude Code & Codex Disk Space on Mac: Safe Cleanup

Claude Code taking up disk space on Mac? First identify the data

Coding agents save more than the app itself. When analyzing developer storage on Mac, a session may include prompts, tool output, file snapshots, logs, and attachments. The project can also accumulate dependencies, generated output, or a worktree. Those are separate kinds of storage, even when they all grew while you were using an AI tool.

The safe first question is not “Which folder can I delete?” It is “Which part is large, what created it, and what would I lose?” Session history can help you resume or review work. Build output may be reproducible. Credentials and configuration may be essential. A folder’s size alone cannot distinguish them.

Claude Code taking up disk space on Mac: where do sessions live?

Claude Code's default user state is ~/.claude. Its project transcripts normally live under ~/.claude/projects/<project>/ as JSONL files. The directory name is derived from the project path, so it may not look like the repository name. You can change the location with CLAUDE_CONFIG_DIR.[1][2]

• projects/ can contain transcripts, tool results, subagent transcripts, and file snapshots used for checkpoint restore.
• history.jsonl keeps prompt history; project memory/ can hold Claude's saved notes.
• debug/, paste-cache/, and other session-related paths may hold logs or intermediate data, depending on features and settings.
• Current session scratchpads and attached images can be under Claude Code's temporary directory, not ~/.claude. On macOS the documented default shape is /private/tmp/claude-<uid>/...; CLAUDE_CODE_TMPDIR can change it.[1]

Anthropic documents a default 30-day retention period for specified Claude Code session data, controlled by cleanupPeriodDays. That does not mean every file under ~/.claude expires: the documentation lists state such as prompt history, memory, credentials, and background-session data with different retention behavior. Use Claude Code's own controls and inspect the current docs for your installed version before changing retention.[1]

Read more: Anthropic: Explore the .claude directory · Anthropic: Manage Claude Code sessions

Where does Codex store its local history and logs on Mac?

Codex stores local state under CODEX_HOME, which defaults to ~/.codex. OpenAI lists configuration, optional file-based authentication, optional history.jsonl, logs, and caches among the contents. Codex's desktop app also documents log files in ~/Library/Logs/com.openai.codex/YYYY/MM/DD, plus active and archived session folders under CODEX_HOME.[3][4]

Treat ~/.codex as an application state folder, not as one disposable cache. auth.json may contain credentials if file-based credential storage is enabled. Configuration can affect how Codex behaves, and transcripts may contain project context or work you want to keep. The app can use a custom CODEX_HOME, so the default path is only a starting point.[3]

OpenAI documents a Codex history setting that can disable local history persistence or cap the history file size. A cap drops older entries; it is a retention choice, not a cleanup command to run casually. Check whether your app, CLI, and IDE integration share the same configured home before assuming one folder represents every Codex component.[3]

Read more: OpenAI: Codex advanced configuration · OpenAI: Codex troubleshooting and log locations

Claude Code taking up disk space on Mac? Measure without deleting

Start with a read-only size check. du reports disk usage; it does not remove or change files. Run it in Terminal while the relevant tools are idle if you want a steadier snapshot, and remember that a missing path or permission error is not evidence that the folder uses zero space.

This checks the default locations only. If you set CLAUDE_CONFIG_DIR or CODEX_HOME, measure those actual paths instead. For a one-level breakdown of one default folder, use du -h -d 1 "$HOME/.claude" | sort -h and repeat with the path you are investigating. macOS may deny access to some directories; do not add sudo just to force a number.

1. In Finder, choose Go > Go to Folder, enter ~/.claude or ~/.codex, and inspect the names and contents. Press Command-Shift-Period to show hidden files in Finder windows.
2. Sort the results by size in a disk-usage tool, then open the path in Finder to identify whether it is a transcript, snapshot, cache, log, credential, or project folder.
3. Check parent folders and modification dates. A large node_modules, build directory, or worktree usually belongs to a project workflow, not to the agent's conversation cache.
4. Keep a note of the measured path and size before taking any action. If the tool is still running, close it normally before using its own cleanup controls.

Read more: Anthropic: Explore the .claude directory · OpenAI: Codex advanced configuration

Which Claude Code or Codex files are safe to remove?

There is no blanket-safe delete rule for either tool's home folder. Prefer each tool's documented retention or cleanup controls for session data. For a clearly identified cache, confirm the owning app's guidance and expected rebuild behavior first. Keep credentials, configuration, memory, transcripts, snapshots, and project files unless you have decided you no longer need them and have a recovery plan.

Anthropic's claude purge can remove a project's transcripts and auto memory along with related state; its documented dry-run shows what it plans to remove. That is intentionally broader than clearing a cache. Read the plan, export anything you want to keep, and do not use a one-line bulk delete against ~/.claude or ~/.codex.[1][2]

Codex's troubleshooting page advises reviewing logs for sensitive information before sharing them. That is another reason not to upload or post a large log folder as a shortcut to diagnosis.[4]

Read more: Anthropic: Explore the .claude directory · Anthropic: Manage Claude Code sessions · OpenAI: Codex troubleshooting and log locations

Why is an application's cache different from session history or project data?

A cache is generally derived data an application can recreate. A transcript records work you may want to resume or review. Project data can include source files, dependencies, uncommitted changes, and local databases. Removing one category can have a very different result from removing another, even when the folders are beside each other.

• A cache may need to be downloaded or generated again, costing time or network data.
• A transcript or prompt history can contain the conversation and tool outputs that help you recover context.
• A file snapshot can be the only local copy of an earlier version that checkpoint restore can use.
• A project dependency folder may be reproducible from a lockfile, but a dirty worktree or generated file may not be.
• A log can include paths, commands, prompts, or other sensitive details, even if it is not project source code.

Read more: Anthropic: Explore the .claude directory · OpenAI: Codex advanced configuration

How can I find unexpected storage from every developer tool?

After checking known Claude Code and Codex locations, scan the whole Mac or the relevant project root. Other tools, package managers, IDEs, simulators, virtual machines, and generated build trees can account for space that appears to be related to an agent simply because it grew during the same work session.

Compare the storage tool's displayed path with the folders you measured. Avoid adding together parent and child sizes, since that double-counts the same data. Also expect differences between logical size and disk allocation, especially for sparse files, shared data, and files you cannot read. If a scan marks a result unknown or incomplete, treat it as unknown rather than zero.

Diskfern can map storage across a selected Mac or folder, then let you inspect large paths in a treemap, list, or sunburst view. Finding and reviewing are separate from removal. Use the map to locate the candidate, then identify its owner and contents before deciding what to keep.