One tool for all your Claude Code conversations.
Zero dependencies. One file. Works everywhere Python 3.7+ runs.
python3 claude-chat.py # Interactive mode
python3 claude-chat.py list # See all your sessions
python3 claude-chat.py search "react hooks" # Search across everything
python3 claude-chat.py export a7e44ed0 --format html # Export with dark theme
python3 claude-chat.py open a7e44ed0 # Render one session straight into the browser
python3 claude-chat.py usage # Monthly token usage table (--by-model to split)
# just want the token table without the tool? one standalone file, zero deps:
# curl -O https://raw.githubusercontent.com/holbizmetrics/claude-code-guide/master/tool/claude_token_usage.py
python3 claude-chat.py serve # Browse in your browser
python3 claude-chat.py wiki --open # Static, searchable HTML archive
python3 claude-chat.py backup --watch # Continuous backup
python3 claude-chat.py protect # Stop auto-deletion
Claude Code stores your conversations as JSONL files in ~/.claude/projects/. Most people don't know this. Worse: Claude Code silently deletes sessions after 30 days by default (the cleanupPeriodDays setting has known bugs where upgrades reset it).
Your conversations contain ideas, decisions, and code you can't reconstruct. This tool makes them searchable, exportable, and safe.
# Just download one file. That's it.
curl -O https://raw.githubusercontent.com/holbizmetrics/claude-code-guide/master/tool/claude-chat.py
# Or grab a tagged release bundle — script + COOKBOOK + docs, zipped:
# https://github.com/holbizmetrics/claude-code-guide/releases/latest
# Or clone the whole repo
git clone https://github.com/holbizmetrics/claude-code-guide.gitRequirements: Python 3.7+. No pip install. No dependencies.
Note: On Windows, use
pythoninstead ofpython3.
Recipe book. Each tip = what to type, what you get, why it matters.
1. First run: stop the 30-day auto-delete.
python3 claude-chat.py protect
Sets cleanupPeriodDays: 99999. One-shot. Do this before anything else.
2. See what you've been working on.
python3 claude-chat.py
list
Enters the REPL, lists recent sessions with [N] numbered refs. Every session gets a short-hash and a summary line.
3. Open any session in your browser — dark theme.
export 1 --format html --open
In the REPL, after list. 1 is the index from the list; you don't have to copy a hash. Works on a fresh REPL too (auto-populates).
4. Turn on math, tables, and links.
export 1 --format html --rich --open
--rich pulls in KaTeX for $...$ math, renders markdown tables as real HTML tables, and makes URLs clickable. Great for research-heavy sessions.
5. See what a session actually did.
export 1 --format html --rich --diagrams --open
--diagrams embeds a mermaid sequence diagram of every tool call (Claude → Read → Bash → Edit …). Zoom, pan, and fullscreen buttons for long sessions. This is the single biggest "oh wow" feature.
6. Find something across all your chats.
python3 claude-chat.py search "sobolev spaces"
python3 claude-chat.py search "auth bug" --project myapp
Full-text search, shows matching snippets with context. Your personal knowledge base you didn't know you had.
7. Pull out just the code.
extract 1 --code
Every code block from the session. Good when you lost a snippet and don't want to re-read 200 messages to find it.
8. Pull out just your own messages.
extract 1 --ideas
Every user message, chronologically. Works like a journal of what you were thinking.
9. Browse visually when you don't know which session you want.
python3 claude-chat.py serve
Opens a local web UI at http://127.0.0.1:3456. Click around sessions, search, read. Localhost only — don't expose it. (Want a permanent, offline, shareable version? wiki --open builds a static searchable HTML archive — same browsing, no server.)
10. Continuous backup in the background.
python3 claude-chat.py backup --watch
Runs as a separate process, polls for changes, keeps the last 5 versions of each session. Zero token cost. Survives Claude Code crashes.
Bonus — in the REPL: prefix any line with ! to run a shell command without leaving. !start file.html on Windows, !open file.html on macOS, !xdg-open file.html on Linux.
Run with no arguments to enter interactive mode:
python3 claude-chat.py
claude-chat v1.0.1 — interactive mode
(prints the full command list: list · search · export · stats · profile ·
compare · activity · extract · serve · wiki · backup · protect,
plus help · quit · !command)
claude-chat> list --limit 5
claude-chat> search "react hooks"
claude-chat> export a7e44ed0 --format html --rich --open
claude-chat> quit
No need to retype python3 claude-chat.py for every command.
For usage from the shell: python3 claude-chat.py help (also works as -h, or
help <command> for a single command's flags). Inside interactive mode, type help.
python3 claude-chat.py list
python3 claude-chat.py list --project crystal # Filter by project
python3 claude-chat.py list --limit 100 # Show more
python3 claude-chat.py list --model fable # Only sessions with that model's turns (substring)
The --model / -m filter (substring match on the model id, e.g. fable → claude-fable-5) is also on search, stats, profile, and activity; compare fable opus diffs two models head-to-head.
Output:
D--MyProject
────────────
a7e44ed0 2026-04-05 23:08 1166KB 0m ago Let's refactor the auth module
c6618ffd 2026-04-05 08:38 819KB 14h ago Fix the deployment pipeline
python3 claude-chat.py search "sobolev spaces"
python3 claude-chat.py search "database migration" --project myapp
Searches all sessions, shows matching snippets with context. Your personal knowledge base you didn't know you had.
python3 claude-chat.py export a7e44ed0 # Markdown (default)
python3 claude-chat.py export a7e44ed0 --format html # Dark-themed HTML
python3 claude-chat.py export a7e44ed0 --format html --open # Export and open
python3 claude-chat.py export a7e44ed0 --format html --rich # Rich: math, tables, links
python3 claude-chat.py export a7e44ed0 --format tex # LaTeX
python3 claude-chat.py export a7e44ed0 --format txt # Plain text
HTML export uses a Tokyo Night dark theme with syntax-highlighted code, collapsible tool calls, and print-friendly CSS.
The --rich flag adds KaTeX math rendering ($...$ and $$...$$), markdown tables as proper HTML tables, clickable links, and markdown headings.
python3 claude-chat.py backup # One-time backup
python3 claude-chat.py backup --watch # Continuous watcher
python3 claude-chat.py backup --watch --interval 5 # Poll every 5s
The --watch mode runs as a separate process, polling for changes. Zero token cost. Survives Claude Code crashes. Keeps the last 5 versions per session.
Backups go to ~/claude-chat-backups/.
python3 claude-chat.py stats
python3 claude-chat.py stats --project myapp
Shows: session count, message counts, tool call frequency, code blocks, models used, largest sessions, project breakdown.
python3 claude-chat.py profile # All models
python3 claude-chat.py profile --model fable # One model (substring match)
python3 claude-chat.py profile --by-session # One row-group per session
python3 claude-chat.py profile --tools # Add per-model tool histogram
Per-model metrics: reasoning %, think-before-action %, first tools reached for, and tool intensity. Add --format json for charting or piping.
python3 claude-chat.py compare fable opus # A vs B, side by side
python3 claude-chat.py compare fable opus --in a7e44ed0 # Within one session (controls for task)
Diffs two models' behavioral profiles (substring match on the model names). --by-session shows A vs B per session; aliased as diff.
python3 claude-chat.py activity # Sessions + turns per day
python3 claude-chat.py activity --by-model # Break the per-day turns down by model
Aliased as timeline. Add --format json for charting.
python3 claude-chat.py extract a7e44ed0 --code # All code blocks
python3 claude-chat.py extract a7e44ed0 --ideas # Your messages only
python3 claude-chat.py extract a7e44ed0 --decisions # Decision points
python3 claude-chat.py serve # Opens http://127.0.0.1:3456
python3 claude-chat.py serve --port 8080 # Custom port
python3 claude-chat.py serve --no-open # Don't auto-open browser
Dark-themed web UI with search, clickable sessions, and full conversation view.
Security note: The server has no authentication and binds to localhost only. Do not forward the port or expose it on a network — your conversations may contain sensitive code and credentials.
python3 claude-chat.py wiki # Build to ~/claude-chat-wiki
python3 claude-chat.py wiki --open # Build and open in browser
python3 claude-chat.py wiki -o ./my-archive # Custom output directory
python3 claude-chat.py wiki --project crystal # Only one project
python3 claude-chat.py wiki --rich # KaTeX math, tables, auto-links
Like serve, but static — no server, nothing to keep running. Produces a folder of self-contained HTML: one cross-linked page per session (8-char session IDs become links to sibling pages, with a "Referenced by" backlinks footer), plus an index.html with built-in client-side search (message bodies and tool-call inputs are indexed). Open index.html straight from disk, search offline, or hand the whole folder to someone else. Rule of thumb: serve to browse live, wiki to keep or share.
python3 claude-chat.py protect
Sets cleanupPeriodDays: 99999 in your Claude Code settings. Run this first if you haven't already.
The Session, Message, and ToolCall classes are importable for building your own tools or GUIs:
import importlib.util
spec = importlib.util.spec_from_file_location("claude_chat", "claude-chat.py")
mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)
sessions = mod.find_all_sessions()
for s in sessions[:5]:
s.parse()
print(f"{s.short_id}: {len(s.messages)} messages — {s.summary()}")
for m in s.user_messages():
print(f" You: {m.text[:80]}")Sessions are JSONL files (one JSON object per line) in:
~/.claude/projects/<project-name>/<session-uuid>.jsonl
Each line contains a message with role, content, tool calls, and model info. The files grow as you talk — a long session can be 20MB+.
Subagents spawned via the Task tool get their own transcripts, nested one level deeper:
~/.claude/projects/<project-name>/<session-uuid>/subagents/agent-<agent-id>.jsonl
list, search, extract, and export all reach these too. You can refer to a subagent by its bare <agent-id> (the same ID printed by Task in the parent session).
For non-obvious patterns — recovering an overwritten file, diagnosing a subagent that errored early, post-compaction session-self-search, anti-fabrication discipline — see tool/COOKBOOK.md.
This repo also hosts The Complete Claude Code Guide — 50+ sections covering installation, hooks, memory, subagents, skills, MCP, YOLO mode, Android/Termux, and more.
Android field guides in android/: the Termux setup script and Disk Forensics with ADB, Shizuku & the force-stop experiment — finding what's silently eating your phone's storage, using Claude Code as the cockpit (real case: a YouTube client bug hoarding 50 GB).
Companion documents:
- Extending Claude Code — field notes from a running installation (July 2026) — the extension surfaces (hooks, skills, transcripts, headless, Monitor, and what's deliberately unused) as wired on a real four-box fleet, with the hard-won lessons and Windows scar tissue.
- Claude Code 2.1.x features & hidden commands
MIT