Stats
Actions
Available In
Tags
Plugin
claude-search
Search your past Claude Code conversations — ranked results, resumable in one command
What's Inside
README
Claude Search
Find any conversation you ever had with Claude Code — in seconds, all local
Ask "where did I work out that postgres pool issue?" and Claude searches every transcript on your machine, ranks the hits, and hands you a claude --resume you can paste straight back. No API key, no upload, no signup — your sessions never leave the laptop.
Quick Start
1. Install the plugin — inside Claude Code, run:
/plugin marketplace add JCodesMore/jcodesmore-plugins
/plugin install claude-search@jcodesmore-plugins
Then fully restart Claude Code (quit the app and reopen).
2. That's it. The first time a session starts, the indexer reads everything under ~/.claude/projects/ once. After that it only re-reads files that changed. No API key, no signup — just start asking.
Try it
Talk to Claude like a friend who remembers everything you've ever done together:
- "Search my sessions for the migration script we wrote last month."
- "Find where I debugged that pgbouncer pool exhaustion."
- "Look up my notes on tokio cancellation safety."
- "What did I figure out about Next.js middleware?"
- "Open a browser UI to browse my history."
Each result comes with a claude --resume <id> command — paste it and you're back in that conversation, exactly where it ended.
What's inside
| Capability | Try saying |
|---|---|
| Ranked full-text search | "Find sessions where I worked on JWT validation" |
| Filter by project | "Search 'auth refactor' in the web-app project" |
| Resume any past session | "Give me a resume command for the pgbouncer one" |
| Local web UI | "Open the search UI in my browser" |
| Show only recent matches | "What did I work on yesterday about Postgres?" |
How it works
- Local-only. Reads
~/.claude/projects/**/*.jsonl(the transcripts Claude Code already writes). Nothing is uploaded. - Incremental. Files unchanged since the last pass are skipped. The first index takes ~30 seconds on a large corpus; later runs are fast.
- Ranked. BM25 with a small user-role boost and recency decay — so the message you wrote yourself ranks above tool output, and recent sessions rank above stale ones.
- Resumable. Every result includes
claude --resume <session-id>. Run it in the original project directory to pick the conversation back up. - Zero deps. Pure Node 20 stdlib. No native bindings, no install step, no npm packages.
Try the CLI directly
If you want to run it outside Claude Code:
node scripts/search.mjs "your query here"
node scripts/search.mjs --web # opens local UI at 127.0.0.1
node scripts/search.mjs --project foo bar # filter by project name
node scripts/search.mjs --json query # machine-readable output
Where the index lives
The index is stored at ~/.claude-search/ (override with CLAUDE_SEARCH_HOME):
index.json— postings offsets, document metadata, session metadatapostings.ndjson— one token per line; loaded on demand for each querydocs.ndjson— one message per line; loaded on demand for snippetsmanifest.json— file signatures so we know what changed
To rebuild from scratch, delete the folder. The next session start (or any /search) will rebuild it.
Advanced install (without the marketplace)
git clone https://github.com/JCodesMore/claude-search.git
cd claude-search
# Use directly from this clone:
claude --plugin-dir ./claude-search
Requirements: Node.js ≥ 20.
Run the tests
npm test
35 tests covering the reader, indexer, searcher, formatter, and web server — all on synthetic JSONL fixtures, no fixtures from your real machine.
License
Apache License 2.0 — © 2026 JCodesMore
Search runs entirely on your local machine. No network calls, no telemetry, no third-party services.
Part of jcodesmore-plugins.
Stats
Version0.2.0
LanguageJavaScript
Stars1
MaintenanceExcellent
LicenseApache-2.0
Last Commit
Added
Available In
Tags
Categories
