Smart Agent for Claude Code

Learn skills from your coding conversations — automatically.

Ever wished Claude Code could learn from your past conversations? That's what we built.

What it does

• 🔍 Watches your Claude Code sessions locally — nothing leaves your machine
• ⚡ Extracts reusable workflow skills AND project knowledge automatically via LLM
• 🔄 Auto-revision — skills get auto-refined as more conversations hit the same pattern
• 🧠 Memory extraction — captures preferences, gotchas, architecture decisions, not just workflows
• 🤝 Team learning — import teammates' session history, every skill shows who contributed
• 📋 Web UI — review, search, filter, approve/reject extracted skills & memories with a side-by-side diff view
• ✅ Writes back approved skills & memories as structured files — CLAUDE.md stays a lean index

Quick Start

npx smart-agent-cc

That's it. On first run:

1. Opens GitHub Device Auth — enter the code in your browser
2. Starts watching ~/.claude/projects/ for new sessions
3. Web UI launches at http://localhost:3456

Architecture

Claude Code Sessions + Teammate Imports
              ▼
         File Watcher (chokidar)
              ▼
        Session Parser (JSONL → conversation)
              ▼
      LLM Analysis (Claude via Copilot API)
          ↙        ↘
      Skills      Memories
      (workflows)  (facts, gotchas, decisions)
          ↘        ↙
     Web UI — Review & Approve
              ▼
      Write to project files:
        .claude/skills/<name>.md
        .claude/memories/<category>.md
        CLAUDE.md (top-10 index)
        ~/.claude/ (global items)
      
     🔄 Auto-Revision Loop

Components

| Component | File | Role | |-----------|------|------| | Watcher | src/watcher.js | Monitors ~/.claude/projects/ for new session files | | Parser | src/parser.js | Parses Claude Code JSONL session files into structured conversations | | Skill Extractor | src/extractor.js | LLM-based extraction of reusable workflow patterns | | Memory Extractor | src/memory-extractor.js | Segmented pipeline: split → extract → conflict check | | Dedup | src/dedup.js | Local tag similarity + LLM confirmation for skill deduplication | | Reviser | src/reviser.js | Auto-revision when a skill is hit 3+ times with new data | | Auth | src/auth.js | GitHub Device Auth Flow — auto token refresh | | Store | src/db.js | SQLite database at ~/.smart-agent-cc/skills.db | | Web UI | web/index.html | Review dashboard with search, filter, sort, diff view | | Writer | src/writer.js | Writes skills/memories to .claude/ and CLAUDE.md | | Import | src/import.js | Import teammate session archives with source attribution |

Output Structure

When skills and memories are approved, they're written as structured files:

your-project/
  CLAUDE.md                        ← Top 10 high-frequency skill summaries (index)
  .claude/
    skills/
      pr_review_triage_and_fix.md  ← One file per approved skill
      figma_to_tailwind_*.md
      ...
    memories/
      gotcha.md                    ← One file per memory category
      project_context.md
      decision.md
      ...

~/.claude/                          ← Global (cross-project) skills & memories
  skills/
  memories/
  CLAUDE.md

CLAUDE.md only contains a lean index with trigger descriptions and file paths — not full content. Claude Code loads it on startup and reads detailed files on demand.

Import teammate sessions

node src/import.js --path ./teammate-sessions/ --project C--src-studio --source "teammate-name" --project-path "C:\src\studio"

Configuration

Optional: create ~/.smart-agent-cc/config.json to override defaults:

{
  "endpoint": "https://api.githubcopilot.com/chat/completions",
  "model": "claude-sonnet-4-20250514"
}

Authentication is handled automatically via GitHub Device Auth — no manual token setup needed.

Memory Categories

| Category | What it captures | |----------|-----------------| | user_preference | Work habits, tool choices, coding style | | factual | Technical facts, configs, API details | | project_context | Architecture, goals, team structure | | gotcha | Bugs, workarounds, unexpected behaviors | | decision | Architecture/design decisions with reasoning | | action_item | Pending tasks, TODOs, follow-ups |

Data Storage

All data is stored locally at ~/.smart-agent-cc/:

• skills.db — SQLite database (sessions, skills, memories, revisions)
• .token — Copilot auth token (auto-refreshed)
• config.json — Optional configuration overrides

License

MIT