claude-smart
A self-improvement plugin for Claude Code, Codex and OpenCode: turns corrections into rules the agent follows in later sessions
Medium risk
We rate an entry medium when the tool runs code, makes network calls or reads project files. Check what exactly it does before installing.
Why this level
- Reads turns, tool calls and replies and stores them locally
- Runs a background service, hooks and a web dashboard on your machine
Install
Manual install
npx claude-smart installRestart Claude Code after installing.
This is third-party code. Review the repository files before installing.
What it does
This is a self-improvement plugin for Claude Code, Codex and OpenCode. Through lifecycle hooks it watches your turns, tool calls and the agent's replies, and turns corrections and working paths into durable rules the agent follows in later sessions. It builds three kinds of artifacts: project-scoped preferences, project-specific skills and shared skills that transfer across projects. Rules do not just accumulate, they get refined: near-duplicates merge and stale ones are archived. Everything runs locally without external calls: semantic search runs on an in-process ONNX embedder, data lives in home directories, and a web dashboard lets you browse history and edit skills.
Who it is for. For developers who want the agent to retain their rules and stop repeating the same mistakes.
Good fit when
- The agent keeps repeating the same mistakes
- You want to lock in project rules and proven paths across sessions
- You want memory and learning kept fully local, without external services
Not a fit when
- You work in Cowork, claude.ai web or a remote environment without local hooks
- A local background service and history storage are undesirable
Example request
Remember that in this repo the local stack starts with pnpm dev:all, not npm run devLimitations
It needs Node.js 20+ and Python 3.12+ plus the host CLI installed. It does not work in Claude Code Cowork, claude.ai web or remote environments without local hooks, since the local backend and home directories are unreachable there. On first run it downloads embedding model weights of about 86 MB once. The plugin reads the contents of your sessions and stores them locally.
How to disable. Run npx claude-smart uninstall (for Codex or OpenCode add --host codex or --host opencode) and restart the host. Data in ~/.reflexio/ and ~/.claude-smart/ is preserved and can be deleted by hand.
Security check
- Reads turns, tool calls and replies and stores them locally
- Runs a background service, hooks and a web dashboard on your machine
README in short
The README describes claude-smart as a self-improvement plugin for Claude Code, Codex and OpenCode that turns corrections and working paths into rules rather than just recording events. It builds project preferences, project skills and shared skills, refining and merging them over time. All data and search are local: an in-process ONNX embedder, hybrid search and storage in home directories. It ships a web dashboard, slash commands and a manual learning marker. Installation is via npx, the license is Apache-2.0, and the reflexio engine is open source.
FAQ
Does data leave my machine?
No. Everything is stored locally and semantic search runs on an in-process ONNX embedder with no external calls.
How is this different from memory?
Memory records what happened. The plugin extracts rules with when-to-apply triggers from corrections and follows them in new sessions.
Related
Open-source personal AI assistant on your own machine: answers in Telegram, Slack, Discord and WhatsApp, extended with skills and plugins
A self-improving agent from Nous Research with a TUI, messaging gateway, cron jobs and skills it writes itself
An open source coding agent for the terminal and desktop with build and plan modes
Open prompt library with a Claude Code plugin, MCP server and CLI: search, fetch and improve prompts and skills from an agent