claude-sqlite-plugin
https://github.com/user-attachments/assets/cb254f18-082b-4fe6-99d9-763d7dcd185f
Embedded conversation viewer and MCP search server for Claude Code. Never lose a conversation, never lose context, even when your network drops mid-session or anthropic suddenly goes bankrupt after IPO😅.
Renamed in v0.2.1. The project was originally claude-postgres-plugin (PostgreSQL-backed). v0.2 swapped the database for embedded SQLite (WAL + FTS5) — no daemon, no createdb, no separate database service — and v0.2.1 renamed everything user-visible to match: plugin id claude-sqlite-plugin, MCP server claude-sqlite, slash commands /csp-*. If you installed v0.2.0 under the old name, run /plugin uninstall claude-postgres-plugin@songtonyli-plugins then reinstall under the new name. The old GitHub URL still redirects.
The problem this solves
You want to copy some conversations from Claude Code to other AI agent to resume the work.
You SSH into a remote dev box and kick off a long Claude Code "vibe coding" session. Two hours in, your home Wi-Fi flakes for 30 seconds. The SSH connection dies. You reconnect, run:
claude --resume
…and Claude immediately autocompacts because the conversation grew large. Half the context — the careful back-and-forth where you nailed down the architecture, the failing test runs, the prompt that finally worked — gets squashed into a one-paragraph summary. The detail is gone.
Same story with: laptop sleep dropping the SSH tunnel, a tmux session that didn't survive a reboot, an iTerm crash, a cellular hotspot blip on the train.
You can take a look at the JSONL file where Claude Code writes to ~/.claude/projects/, and you will be surprised that not every conversation, tool result, or assistant answers are saved! But by default, you have no good way to search it, browse it, or hand the relevant slice back to a fresh Claude session — so --resume and its autocompact are your only options.
This plugin fixes that:
- A real-time watcher ingests every message into a local SQLite database the moment Claude Code writes it to disk. ACID-safe via WAL +
synchronous = FULL + foreign keys. No transaction is ever lost, even if the parent terminal dies.
- A web dashboard at
http://localhost:3456 lets you browse, search, and export every past session — text, tool calls, thinking blocks, and image / document attachments rendered exactly as they appeared.
- An MCP server ships with the plugin so Claude itself can search your conversation history during a new session: "what did we try for the rate-limiter bug last week?" → Claude calls the
search_messages tool → answers with real evidence from your past work.
How it works
You use claude normally This plugin runs in background
| |
v v
claude-code writes fs.watch detects
~/.claude/projects/*.jsonl --> every changes in CLI
|
v
Parser extracts messages,
tool calls, thinking, images
|
v
SQLite stores everything
(WAL + FTS5, ACID safe)
|
+-----------------------------+
| |
v v
Web dashboard at :3456 MCP tools for Claude
(humans browse visually) (Claude searches via
search_messages, etc.)
Two ways to install
Option A — As a Claude Code plugin (recommended)
This bundles the MCP server, the slash commands, and the watcher all in one install.
/plugin marketplace add SongTonyLi/claude-sqlite-plugin
/plugin install claude-sqlite-plugin@songtonyli-plugins
Prerequisites: Bun installed (curl -fsSL https://bun.sh/install | bash). That's it — no Postgres, no createdb, no database service to manage. The DB is a single SQLite file under ~/.claude-sqlite-plugin/csp.sqlite (or ${CLAUDE_PLUGIN_DATA} if Claude Code provides one).
Dependencies are installed automatically on the first MCP server start — no manual bun install required. If you also want the web dashboard, build the frontend once:
# Replace the path below with whatever /plugin install reported, typically:
cd ~/.claude/plugins/cache/songtonyli-plugins/claude-sqlite-plugin/0.2.3
(cd web && bun install && bun --bun vite build) # frontend bundle (only needed if you'll use the dashboard)
