You have a dozen Claude Code sessions running across different git worktrees. Some are half-finished, others you won't touch for weeks. Keeping them open wastes resources; closing them makes it harder to pick up where you left off.

claude-park lets you shelve a Claude Code session and resume it whenever you're ready. It preserves your conversation, protects your worktree, and generates a concise summary of where you left off.

Two commands:

!park     # inside Claude Code: park this session and close its tab
parked    # in any shell: find a parked session and resume it

Install

./install.sh

This links claude-park, park, and parked into ~/.local/bin.

Requirements: macOS, Python 3.9+, and git. iTerm2 is needed to automatically close and reopen tabs; without it, parking still stops Claude but leaves the tab open.

Optional dependencies:

• fzf for the interactive session picker
• gh for GitHub pull request information

Optional: /park skill

If your Claude session is still warm, the optional /park skill can write its own summary using the context already available to the model, avoiding a separate Haiku summarization call.

claude plugin marketplace add hotdata-dev/claude-park
claude plugin install park@claude-park

Why park a session?

Free up your workspace. Stop idle Claude processes, close unused terminal tabs, and reclaim disk space from rebuildable dependencies.

Protect unfinished work. Lock your git worktree against accidental removal and leave a PARKED.md notice explaining that the work is parked, not abandoned.

Never lose your place. Preserve the full conversation and generate a concise park card capturing the goal, progress, next steps, and gotchas.

Resume without starting over. Return to the original conversation, or start a fresh session using the park card instead of re-sending the entire history.

How parking works

Park without re-sending your context

!park is a shell command, not a Claude slash command. Parking doesn't invoke your active Claude model or re-send its entire conversation.

Instead, a background job passes a trimmed transcript excerpt (opening request, latest compaction summary, and recent turns, up to roughly 60K characters) to Haiku, which generates a park card containing:

• The task's goal and current progress
• What has already been completed
• Next steps
• Important context and gotchas

The result is saved alongside the session's metadata, ready whenever you return.

Protect your work

Parking does more than save the conversation:

1. Records session state. Saves the title, directory, branch, ahead/dirty counts, last commit, associated PR (via gh), and last prompt in ~/.claude/parked/<session-id>.json.
2. Locks the worktree. Uses git worktree lock with a claude-park: reason. Git refuses to remove or prune locked worktrees. Existing locks are preserved.
3. Leaves a notice. Writes PARKED.md at the worktree root with the resume command and park card. The file is hidden from git status via .git/info/exclude. An existing PARKED.md not created by claude-park is never overwritten.
4. Reclaims disk space. Removes gitignored, rebuildable directories such as target, node_modules, .venv, and .next. Only paths reported as ignored by git are touched. Use --no-clean to skip cleanup.
5. Preserves the transcript. Hard-links the conversation into ~/.claude/parked/transcripts/, protecting it from Claude Code's automatic transcript cleanup without duplicating its disk usage. The original is relinked on resume if necessary.
6. Closes the session. Stops the Claude process and closes its iTerm2 pane.

Resuming a session

Run parked from any shell to browse your parked sessions.

The interactive fzf picker shows each session's age, repository, title, progress, and git state:

  3d  api-server  Rate limiter for the public search API     ▰▰▰▰▱ review  PR #482 open
  9d  web-app     Migrate settings page to new form library  ▰▰▱▱▱ design  2 uncommitted

A lightweight Haiku pass over the park card generates a short preview with:

• A descriptive title
• Current phase (research → design → build → review → wrap-up)
• One-line status
• Top three next steps

Resume the full conversation, or start fresh

You have two ways to pick up your work.

Resume the original session to restore the full conversation exactly where you left it:

parked resume <id>

Start fresh with the park card to give a new Claude session the essential context without re-sending the entire conversation:

parked resume --fresh <id>

The fresh option is especially useful for long-running tasks where the original conversation has accumulated hundreds of thousands of tokens.

Picker shortcuts

Key	Action
enter	Resume the original session in the current terminal
ctrl-t	Resume in a new iTerm2 tab
ctrl-f	Start a fresh session seeded with the park card
ctrl-x	Drop the parked session without resuming
ctrl-o	Toggle between the short preview and full park card

Resuming or dropping a session unlocks its worktree, removes the generated PARKED.md, and archives the park card.

Park and resume repeatedly

You can park the same session more than once without losing context.

When you park it again, claude-park starts from the archived card:

• No new conversation turns? The existing card is reused without another model call.
• New conversation turns? Haiku updates the card using only the new turns, preserving older context captured in previous summaries.
• Summary update fails? The previous card is retained and marked stale, and the next park retries.

This allows context to survive repeated parking and resuming, even as older conversation history moves outside the summarization window.

Token usage

claude-park is designed to minimize unnecessary model calls and context reprocessing.

Action	Token usage
Park with !park	A Haiku call summarizes a trimmed transcript excerpt; the full conversation is not re-sent to your active Claude model
Park using /park	Uses the active session's existing context to generate the card, avoiding a separate summarization call; a small Haiku call still turns the card into the picker preview
Resume original session	Claude may need to re-send the full conversation history on the next turn
Resume fresh session	Starts with the much smaller park card instead of the full conversation
Park again with no new turns	Reuses the existing card without a model call

The full conversation remains available when you need it. The fresh-session option gives you a way to continue without paying to reconstruct all of that context.

Command reference

Everything available in the interactive picker also works directly from the command line.

# Browse and inspect parked sessions
parked list
parked show [--brief] [--color] <id-prefix|title>

# Resume a session
parked resume [--tab] [--fresh] <id-prefix|title>

# Remove a session from parking without resuming
parked drop <id-prefix|title>

# Refresh generated summaries and previews
parked refresh [--all] [<id>]
parked summarize <id-prefix|title>

# Check transcript compatibility
parked doctor [--session <id>]

# Customize parking behavior
park --no-clean | --no-close | --no-summary | --no-pr

Configuration

Configure claude-park using environment variables:

Variable	Description	Default
CLAUDE_PARK_MODEL	Model used for generating park cards	haiku
CLAUDE_PARK_DIR	Directory where parked sessions are stored	~/.claude/parked

Claude Code compatibility

claude-park reads Claude Code's local transcripts (~/.claude/projects/*/<session-id>.jsonl) and the CLAUDE_CODE_SESSION_ID environment variable.

These are not documented, stable interfaces. A future Claude Code update could change their format.

Rather than silently producing empty or incorrect summaries, claude-park fails loudly when it encounters an unrecognized transcript format:

• If a transcript contains entries but no readable turns, parking skips summarization and reports the incompatibility.
• Each park card records the Claude Code version that produced the transcript.
• parked doctor inspects your latest transcript (or a specified session) and reports which fields it recognizes.

After upgrading Claude Code, you can check compatibility with:

parked doctor

The command exits non-zero if the transcript format is not recognized.

Development and tests

Run the test suite with:

python3 -m unittest discover -s tests

The demo shown above uses fictional data in /tmp/claude-park-demo, with stand-ins for claude and gh.

To regenerate it:

docs/demo/render.sh

Requires VHS, fzf, ffmpeg, and Pillow.

License

MIT. See LICENSE.

claude-park is an independent project by Hotdata, Inc. for Claude Code. It is not affiliated with or endorsed by Anthropic.