sessionbook

Save Claude Code sessions as self-contained HTML files.

sessionbook wraps the claude CLI and, on exit, converts the session's JSONL log into a .html file in the local .sessionbook/ directory. The output is a chat-style transcript with collapsible thinking blocks, user-choice decision cards, and linked sub-agent transcripts. Each file is self-contained (inline CSS, no external dependencies) and opens in any browser.

Requirements

• Python 3.10+
• Claude Code CLI installed and on $PATH
• macOS (uses os.fork)

Installation

pip install sessionbook

For development (includes pytest):

pip install -e ".[dev]"

Quick Start

Capture a session

sessionbook claude

This launches Claude Code exactly as if you ran claude directly. All interactive features work unchanged (Ctrl+G, slash commands, MCP tools). When the session ends, sessionbook reads the JSONL session file and saves an HTML file to .sessionbook/<TIMESTAMP>.html.

If you use /clear during a session, each sub-session is saved as a separate file.

Arguments are forwarded

Any arguments after claude are passed through to the Claude Code CLI:

sessionbook claude --model sonnet
sessionbook claude -p "Explain this repo"

Convert existing sessions

Retroactively convert past Claude Code sessions for the current project directory:

sessionbook sync

Or convert a specific session by ID:

sessionbook sync <session-id>

Output Format

HTML files are written to .sessionbook/ in the working directory. Each file is named with a local-time timestamp: 2026-02-07T14-32-01.html.

.sessionbook/
  2026-02-10T21-43-08.html
  2026-02-10T21-43-08/          # sub-agent transcripts (if any)
    agent-aa5abe3.html
    agent-bc7f012.html

Session layout

Each turn in the conversation is rendered as a visually distinct card:

• User turns — light blue background with the user's prompt text
• Assistant turns — light gray background with the model's response
• Thinking blocks — collapsed <details> elements (click to expand)
• Decision cards — when the model presented options via AskUserQuestion, the card shows the question, all options, and which one was selected
• Sub-agent cards — when the model forked work to a sub-agent (e.g., Explore, executor), the card shows a summary and links to a separate HTML file with the full sub-agent transcript

What is captured

Content	Included
User prompts	Yes
Model responses (text)	Yes
Thinking blocks	Yes (collapsed)
User choice interactions	Yes (decision cards)
Sub-agent transcripts	Yes (linked out)
Tool use / tool results	No (filtered out)
System context (isMeta)	No (filtered out)

CLI Reference

sessionbook <subcommand> [options] [args...]

Subcommands:
  claude    Wrap Claude Code and save session as HTML on exit
  sync      Convert existing Claude Code sessions to HTML

Options:
  --verbose    Enable debug-level logging to stderr
  --version    Print version and exit
  --help       Print usage and exit

Environment variables:
  sessionbook_DEBUG=1    Equivalent to --verbose

Exit codes mirror the claude child process. If claude exits 0, sessionbook exits 0.

How It Works

1. sessionbook claude calls os.fork() + os.execvp("claude", ...). The child process inherits the terminal directly, so Claude Code behaves identically to a direct invocation. The parent process is blocked in os.waitpid() and never touches stdin/stdout.
2. After the child exits, the parent reads JSONL session files from ~/.claude/projects/ that were modified during the session.
3. Each session's user and assistant turns are extracted (with requestId collapsing for streamed responses), along with thinking blocks, user choices, and sub-agent references.
4. The session is rendered as self-contained HTML with inline CSS and written atomically (tempfile + rename) to .sessionbook/.

Zero runtime dependencies. The HTML is constructed from Python f-strings with html.escape() for all user content.

Uninstall

pip uninstall sessionbook

The .sessionbook/ directories contain only generated HTML files and can be deleted at your discretion.

License

MIT