Session Browser

Browse and search your Claude Code conversation history in a local web UI.

Features

• Session list with sortable columns: date, repos touched, first prompt, branch, duration, cost, turns
• Full-text search across all user messages and assistant responses (SQLite FTS5)
• Filter by workspace/repo, date range (today / 7d / 30d / all)
• Session detail view with chat-style conversation rendering:
  • Markdown rendering with syntax-highlighted code blocks
  • Collapsible thinking blocks
  • Collapsible tool calls with one-line summaries (Bash commands, file operations, agent dispatches)
  • Tool results with large output truncation
  • Subagent transcript navigation
• Auto-indexing with incremental re-scan (only re-indexes changed files)
• Repo detection from file paths in tool calls (auto-discovers git repos)
• Dark/light theme toggle
• No LLM involved -- pure data extraction and rendering

How it works

Claude Code stores conversation transcripts as JSONL files in ~/.claude/projects/. Session Browser indexes these files into a local SQLite database and serves a web UI for browsing them.

~/.claude/projects/**/*.jsonl  ->  [Indexer]  ->  SQLite  ->  [FastAPI]  ->  [Browser]

Install

Requires Python 3.11+.

git clone https://github.com/giannimassi/session-browser.git
cd session-browser
python3 -m venv .venv
.venv/bin/pip install -e .

Usage

.venv/bin/python3 server.py

Open http://127.0.0.1:8787 in your browser.

On first launch, all sessions are indexed (takes 2-5 seconds for ~300 sessions). Subsequent launches only re-index new or modified files.

Data

Session Browser is read-only. It never modifies your Claude Code data. The SQLite index is stored at ~/.cache/session-browser/sessions.db and can be safely deleted (it will be rebuilt on next launch).

What gets indexed

Per session:

• Session ID, project, git branch, timestamps, duration
• First user prompt
• Token usage and estimated cost (Opus pricing)
• Tool call counts
• Repos touched (auto-detected from file paths in Read/Write/Edit/Bash tool calls)
• Full text of user messages and assistant responses (for search)

What gets parsed on detail view

• Complete conversation with user/assistant messages
• Thinking blocks (collapsed by default)
• Tool calls with inputs and results
• Subagent transcripts (linked from Agent tool calls)

Tech stack

• Python -- FastAPI + uvicorn
• SQLite with FTS5 -- session index and full-text search
• Vanilla JS + CSS -- no build step, no node_modules
• marked.js + highlight.js -- vendored for markdown/code rendering

Configuration

The server binds to 127.0.0.1:8787 by default. The Claude data directory defaults to ~/.claude. Both can be changed in server.py.

License

MIT