Your AI that trains itself. MCP-native self-improving agent for Claude Code.
Every correction becomes a rule. Every rule makes Claude smarter. Automatically.
- Contributing: see CONTRIBUTING.md
- Security: see SECURITY.md
/plugin install claude-learner@unisone/claude-learnernpm install -g claude-learner
claude-learner initThat's it. Claude now learns from every session.
You correct Claude. Claude forgets. You correct again.
You: "Don't use rm, use trash"
Claude: *uses rm again next session*
You: π€¦
Knowledge dies when the session ends.
claude-learner watches your sessions, detects patterns, and creates permanent rules:
[daemon] Detected: User corrected "rm" β "trash" (3 times)
[daemon] π Proposed rule: "Use trash instead of rm"
[you] claude-learner approve rule_xxx
[claude] *follows rule forever*
π You work with Claude Code
β
ποΈ Daemon watches sessions in real-time
β
π Detects patterns (corrections, retries, rollbacks)
β
π Proposes rules for your approval
β
β
Approved rules become permanent
β
π― Claude follows them via MCP
β
π Ineffective rules get auto-pruned
The daemon runs in the background. You don't need to do anything except approve good rules.
| Command | What it does |
|---|---|
init |
One-step setup (starts daemon + registers MCP) |
start |
Start the learning daemon |
stop |
Stop the daemon |
status |
Show daemon status + stats |
watch |
Live activity feed |
rules |
List all rules |
rules --pending |
Show rules awaiting approval |
rules --effectiveness |
Show compliance rates |
approve <id> |
Approve a proposed rule |
reject <id> |
Reject a proposed rule |
mcp-serve |
Start MCP server (for Claude Code) |
| analyze | Batch-analyze session history |
| improve | Generate CLAUDE.md suggestions |
| export | Export learnings to files |
| stats | Usage statistics |
When integrated with Claude Code, these tools are available:
| Tool | Purpose |
|---|---|
get_rules |
Load active rules at session start |
check_rule |
Check if action violates a rule |
log_correction |
Log corrections for learning |
record_compliance |
Track rule effectiveness |
get_pending_rules |
View proposed rules |
approve_rule / reject_rule |
Manage rules |
Claude Code automatically calls these to learn and improve.
- You correct Claude: "Don't use
any, use proper types" - Daemon detects it: Logs as correction pattern
- Pattern repeats: Same correction 2+ times
- Rule proposed:
"Don't use any type" - You approve:
claude-learner approve rule_xxx - Rule active: Claude checks it before using
any - Tracked: System monitors compliance
- Auto-prune: If ignored >70%, rule is pruned
| Scope | Applies to | Example |
|---|---|---|
global |
All projects | "Use 2-space indentation" |
project |
Specific project | "This repo uses pnpm" |
file |
File pattern | "*.test.ts files use vitest" |
~/.claude-learner/
βββ learner.db # SQLite database (rules, patterns, sessions)
/tmp/
βββ claude-learner.pid # Daemon PID
βββ claude-learner.log # Daemon logs
βββββββββββββββββββ ββββββββββββββββ βββββββββββββββ
β Claude Code ββββββΆβ MCP Server ββββββΆβ SQLite β
β (your work) β β (7 tools) β β (storage) β
βββββββββββββββββββ ββββββββββββββββ βββββββββββββββ
β β²
βΌ β
βββββββββββββββββββ ββββββββββββββββ β
β Session Files ββββββΆβ Daemon βββββββββββββ
β ~/.claude/... β β (watcher) β
βββββββββββββββββββ ββββββββββββββββ
- Node.js 20+
- Claude Code (for MCP integration)
Does it send data anywhere?
No. Everything runs locally. Your sessions never leave your machine.
Can I use it without the daemon?
Yes. Use analyze/improve/export for batch processing.
How do I uninstall?
claude-learner stop
npm uninstall -g claude-learner
rm -rf ~/.claude-learnerIf this saved you time:
- β Buy Me a Coffee
- π GitHub Sponsors
- β Star the repo
- π Changelog
- π Releases
- π Issues
- π‘ Discussions
MIT Β© Alex Zaytsev
Make Claude Code actually learn. Install in 30 seconds. Never repeat a correction.
- intel-skill β Market intelligence for Claude Code (
npx skills add unisone/intel-skill) - moltbot-config β Production memory engine, self-review system, and configs for Moltbot
- agentpulse β Real-time agent monitoring dashboard