Claude OS

Give Your AI a Memory
Claude Code that actually remembers you.

---

✨ The Magic

You: "Remember this: our API uses JWT tokens with 15-minute expiry and 7-day refresh tokens"

Claude: Saved: "API Authentication Strategy" (Architecture)

That's it. Next week, next month, next project - Claude remembers. No commands to memorize. No complex setup. Just talk naturally.

"remember this..."     → saved to your knowledge base
"what did we decide?"  → searches your memories
"how did we fix that?" → finds past solutions

Every conversation makes Claude smarter. Every insight builds your shared knowledge.

---

🆕 What's New in v2.5

Latest Release: February 2026

🔍 Cross-KB Search

Search across all your knowledge bases at once with a single query!

mcp__code-forge__search_all_knowledge_bases
  query: "authentication patterns"
  kb_filter: "MyProject-"          # optional: limit to one project's KBs

• Results merged by relevance score with KB attribution
• Automatic deduplication across KBs
• Optional prefix filter to scope by project
• New API endpoint: POST /api/kb/search-all

🩺 Inline Health Checks

Health checks now run automatically during search — no manual invocation needed.

• When you search a KB, Claude OS checks if a health report has been run in the last 24 hours
• If stale, runs a quick health check and appends HIGH/CRITICAL warnings to the search result
• Cached in data/health_cache.json — never slows down or breaks search
• Look for _health_warnings in search results

📝 Simplified Session Management

Session state reduced from a 50-field JSON blob to 4 fields:

{
  "last_task": "Fix appointment email flood",
  "last_branch": "fix-email-flood",
  "stopped_at": "2026-02-06T18:30:00Z",
  "one_liner": "Fixed dedup check, still need rate limiting"
}

• START: read state, show one-liner, search memories, ready
• END: git diff summary, offer to save, write state
• save, blocker, pattern sub-commands unchanged

📄 Leaner CLAUDE.md Template

Project CLAUDE.md template cut from 351 to 128 lines:

• Removed 200-line mandatory session protocol (6-phase startup, ASCII prompts)
• Replaced with 4-line "Session Tips" section
• Project content first, Claude OS section second
• All template variables preserved

📋 Recent Improvements

Version	Highlights
v2.5	Cross-KB search, inline health checks, simplified sessions, leaner templates
v2.4	Knowledge lifecycle engine (dedup, consolidate, archive, health)
v2.3	Skills library, community skills, session insights
v2.2	Gum CLI support, safety features, lite model default
v2.1	Unified installer, OpenAI provider support
v2.0	Hybrid tree-sitter indexing, real-time Kanban board

🙌 Community Contributions

Thanks to our amazing contributors!

PR	Contributor	Description
#22	@illAssad	Fix delete document endpoint SQLite cursor handling
#21	@illAssad	Skip node_modules and build directories during ingestion
#20	@illAssad	Fix tree-sitter version compatibility
#19	@illAssad	Add non-blocking semantic indexing with Jobs Dashboard UI
#18	@illAssad	Fix frontend startup by auto-installing npm dependencies
#17	@williamclavier	Fix: Ensure commands/skills directories exist
#16	@jplimack	Fix hardcoded paths - make dynamic
#12	@gkastanis	Add missing frontend lib files
#11	@gkastanis	Add Linux support for installation
#10	@nicseltzer	Fix broken README link

---

🚀 What is Claude OS?

Claude OS isn't just a tool - it's Claude's memory.

Think about it: you and Claude work together on a feature. You explain your architecture, your patterns, your preferences. Then you close the terminal... and it's all gone. Tomorrow, you start over.

What if Claude actually remembered?

Before Claude OS

Day 1: "We use JWT tokens with refresh..."
Day 2: "As I mentioned, we use JWT tokens..."
Day 3: "Again, the auth system uses JWT..."

After Claude OS

Day 1: "Remember this: we use JWT tokens with refresh..."
Day 2: Claude already knows. Applies the pattern automatically.
Day 3: Claude suggests improvements based on what worked.

The Difference

Without Claude OS	With Claude OS
Explain the same things repeatedly	Claude remembers your decisions
Start cold every session	Context loaded automatically
Patterns forgotten	Patterns compound over time
Claude is a tool	Claude is a partner

Why It Works

• 🧠 Persistent Knowledge - Decisions, patterns, solutions saved across sessions
• 🔍 Automatic Recall - Relevant memories surface when you need them
• 📚 Documentation Search - Your docs indexed and searchable via RAG
• 🎯 Session Learning - Claude extracts insights from past conversations
• 🔒 100% Local - Your knowledge never leaves your machine

Features at a Glance

Feature	What It Does
Natural Language Memory	Just say "remember this" - no commands needed
Session Insights	Auto-extracts patterns from past conversations
Lightning Indexing	10,000 files indexed in 30 seconds
Skills Library	36+ community skills, one-click install
Cross-Project Learning	Patterns from Project A help in Project B
Knowledge Lifecycle	Dedup, consolidate, archive, and health reports
One-Command Setup	/claude-os-init and you're ready

---

⚡ NEW: Hybrid Indexing System

Claude OS v2.0 introduces lightning-fast tree-sitter based indexing!

The Problem with Traditional Indexing

Previous versions embedded EVERY file, which was painfully slow for large codebases:

• Large projects (10,000+ files): 3-5 hours to index
• Must complete before Claude can start working
• High resource usage, blocks productive coding

The Solution: Hybrid Two-Phase Indexing

Inspired by Aider's approach, Claude OS now uses:

Phase 1: Structural Index (30 seconds)

• ⚡ Parse files with tree-sitter (no LLM calls!)
• 📊 Extract symbols only (classes, functions, signatures)
• 🔗 Build dependency graph
• 🏆 PageRank importance scoring
• ✅ Ready to code immediately!

Phase 2: Semantic Index (optional, background)

• 🎯 Selective embedding (top 20% most important files)
• 📚 Full embedding for documentation
• 🔍 Deep semantic search when needed
• ⏰ Runs in background while you code

Performance Comparison

Feature	Before	After (Hybrid)
Large project (10k files)	3-5 hours	30 seconds + 20 min optional
Files embedded	100,000+ chunks	~20,000 chunks (80% reduction)
Start coding	After full index	Immediately!
Resource usage	High Ollama load	Minimal CPU/memory
Query speed	Semantic search	Instant structural + semantic

📖 Read the full design: docs/HYBRID_INDEXING_DESIGN.md

---

🎨 Visual Interface

Claude OS provides a beautiful, intuitive web interface for managing your AI development workflow:

Welcome Screen Get started with Claude OS	Project Overview View project details and MCP status
Kanban Board Track spec implementation progress	Services Dashboard Monitor all Claude OS services

📖 See the complete visual guide: docs/guides/VISUAL_GUIDE.md

---

🏗️ Architecture Overview

Claude OS is built on 5 core pillars that work together to give Claude persistent memory:

1. 🧠 Real-Time Learning - Automatically captures insights from conversations via Redis Pub/Sub
2. 💾 Memory MCP - Persistent memory system with instant recall using natural language
3. 🔍 Analyze-Project - Intelligent codebase indexing with git hooks and tree-sitter
4. 🎯 Session Management - Auto-resume sessions with full context preservation
5. 📚 Semantic Search - Vector-based code understanding and pattern recognition

All knowledge flows through the Semantic Knowledge Base (SQLite + sqlite-vec), exposed via the MCP Server (port 8051) to Claude Code, giving you an AI assistant that never forgets.

Data Flow: Git Commit → 3s indexing → SQLite → MCP → Claude → You

---

💻 Installation & Setup

Prerequisites

Required:

• macOS or Linux (Ubuntu, Debian, Fedora, RHEL, Arch)
• Python 3.11 or 3.12 (python3 --version)
  • Note: Python 3.13+ not yet supported due to dependency constraints
• Git (git --version)

Optional:

• Node.js 16+ (for React UI)
• Ollama (for local AI) or OpenAI API key

Note: Windows support coming soon.

Quick Installation

# Clone the repository
git clone https://github.com/brobertsaz/claude-os.git
cd claude-os

# Run the unified installer
./setup-claude-os.sh

First time? Try the demo first:

./setup-claude-os.sh --demo    # See the beautiful UI (no changes made)
./setup-claude-os.sh --dry-run # Preview what would be installed

The installer will guide you through:

1. Choose Provider: Local (Ollama) or Cloud (OpenAI)
2. Choose Model Size: Lite (2GB) or Full (4.7GB) - for local installs
3. Automatic Setup: Python, dependencies, MCP server, commands/skills

What gets installed:

• ✅ Python virtual environment
• ✅ All dependencies
• ✅ MCP server configuration
• ✅ Commands and skills symlinked to ~/.claude/
• ✅ Ollama + models (if local provider selected)
• ✅ Redis for caching

Installer Options

./setup-claude-os.sh           # Interactive installation
./setup-claude-os.sh --demo    # Try the UI without changes
./setup-claude-os.sh --dry-run # Preview what would happen
./setup-claude-os.sh --help    # Show all options
./setup-claude-os.sh --version # Show version

Legacy Scripts

The old install.sh and setup.sh scripts still work - they redirect to the new unified installer.

Visit http://localhost:5173 to use the web UI.

Starting Claude OS

After installation, start the services:

./start.sh

This starts the MCP server at http://localhost:8051

---

🚀 Quick Start

Initialize any project with Claude OS in under 2 minutes:

Step 1: Navigate to Your Project

cd /path/to/your/project

Step 2: Initialize with Claude OS

In Claude Code, run:

/claude-os-init

The command will:

1. Ask Questions Interactively:

   • Project name (auto-detects from folder)
   • Tech stack (Ruby on Rails, Python, Node.js, etc.)
   • Database (PostgreSQL, MySQL, etc.)
   • Development environment (Docker, Local, etc.)
   • Brief description
   • Documentation directory to ingest (optional)

2. Create Project in Claude OS:

   • Calls API to create project
   • Creates 4 knowledge bases automatically:
     • {project}-project_memories - Claude's memory
     • {project}-project_profile - Architecture & standards
     • {project}-project_index - Codebase index
     • {project}-knowledge_docs - Your documentation

3. Set Up Project Structure:

   your-project/
   ├── CLAUDE.md           # Auto-loaded every session!
   ├── .claude/            # Commands, skills, agents
   │   ├── ARCHITECTURE.md
   │   ├── CODING_STANDARDS.md
   │   └── DEVELOPMENT_PRACTICES.md
   └── .claude-os/         # Config and state (git-ignored)
       ├── config.json
       └── hooks.json
   

4. Ingest Documentation:

   • Scans your docs directory
   • Uploads all files to {project}-knowledge_docs
   • Creates vector embeddings for search

5. Analyze Codebase:

   • Runs initialize-project skill
   • Generates coding standards
   • Documents architecture
   • Indexes key files

6. Ready to Code:

   • Claude now knows your project
   • Memory persists across sessions
   • Context auto-loads on session start

What You Get

• ✅ 4 knowledge bases created (memories, profile, index, docs)
• ✅ Documentation auto-indexed
• ✅ Codebase analyzed
• ✅ CLAUDE.md file with all context
• ✅ Ready to code with AI memory!

---

🧠 How Claude OS Works

Session Workflow

Every Claude Code session automatically:

1. Checks for Active Session

   • Reads claude-os-state.json
   • Prompts: Continue working? Start something new?

2. Loads Context

   • Searches {project}-project_memories for recent work
   • Loads relevant patterns and decisions
   • Shows what it remembers

3. Works With Memory

   • Saves insights with /claude-os-remember
   • Searches memories with /claude-os-search
   • References past decisions automatically

4. Ends Session

   • Saves session summary
   • Updates memories
   • Tracks what was accomplished

Available Commands

All these work in any initialized project:

• /claude-os-init - Initialize new project
• /claude-os-search [query] - Search memories & docs
• /claude-os-remember [content] - Quick save to memories
• /claude-os-save [title] - Full-featured save with KB selection
• /claude-os-list - List all knowledge bases
• /claude-os-session [action] - Manage development sessions
• /claude-os-triggers - Manage trigger phrases
• /claude-os-skills [action] - Manage skills (list, install, create)
• /claude-os-lifecycle [action] - KB health, dedup, consolidate, archive

Available Skills

Global Skills (always available):

• initialize-project - Analyze codebase and generate standards
• memory - Save and recall information (supports "remember this:", "save to memory", etc.)

Community Skills (install via /claude-os-skills):

• 36+ skills from Anthropic Official and Superpowers repos
• PDF/XLSX manipulation, frontend design, TDD, debugging, code review, and more

---

🤖 Agent-OS: Spec-Driven Development (Optional)

Created by Builder Methods (CasJam Media LLC) MIT Licensed • Separate Optional Integration

Agent-OS adds structured workflows for planning and implementing features using 8 specialized agents.

Agent-OS is a separate open-source project that can be installed alongside Claude OS. We're grateful to Builder Methods for creating such powerful spec-driven development tools.

Manual Installation

If the Agent-OS repository is available, you can install it with:

git clone https://github.com/buildermethods/agent-os.git ~/.claude/agents/agent-os

Note: Check if the repository exists before attempting to install.

When to Use Agent-OS

If you have Agent-OS installed, use it when you want:

• Structured feature planning with iterative requirements gathering
• Detailed specifications before coding
• Task breakdowns with clear implementation steps
• Verification workflows to ensure completeness

The 8 Agents

Specification Workflow:

1. spec-initializer - Initialize new spec directories
2. spec-shaper - Gather requirements through 1-3 questions at a time
3. spec-writer - Create detailed technical specifications
4. tasks-list-creator - Break specs into actionable tasks

Implementation Workflow:

5. implementer - Implement features following task list
6. implementation-verifier - Verify implementation completeness
7. spec-verifier - Verify specs and tasks consistency
8. product-planner - Create product documentation

Agent-OS Commands

Available when enabled:

• /new-spec - Initialize a new feature specification
• /create-spec - Full specification workflow (gather requirements → create spec → generate tasks)
• /plan-product - Create product mission, roadmap, and tech stack docs
• /implement-spec - Implement a specification following its tasks

How It Works

1. User: "/new-spec user-authentication"
   → Agent creates spec directory structure

2. User: "/create-spec"
   → spec-shaper asks 1-3 questions at a time
   → Gathers requirements iteratively
   → Identifies reusable code
   → Collects visual assets

3. Agent: spec-writer creates detailed specification
   → tasks-list-creator generates actionable tasks

4. User: "/implement-spec user-authentication"
   → implementer follows tasks step-by-step
   → implementation-verifier checks completeness

5. Result: Fully specified, implemented, and verified feature!

Agent-OS Project Structure

When enabled, your project gets:

your-project/
├── agent-os/
│   ├── config.yml          # Agent-OS configuration
│   ├── product/            # Product documentation
│   │   ├── mission.md      # Product mission
│   │   ├── roadmap.md      # Feature roadmap
│   │   └── tech-stack.md   # Technology stack
│   ├── specs/              # Feature specifications
│   │   └── YYYY-MM-DD-feature-name/
│   │       ├── planning/
│   │       │   ├── requirements.md
│   │       │   └── visuals/
│   │       ├── spec.md
│   │       └── tasks.md
│   └── standards/          # Coding standards (as skills)
└── .claude/agents/agent-os/  # 8 agents (symlinked)

Integration with Claude OS

Agent-OS agents deeply integrate with Claude OS:

• Search memories before creating specs (avoid reinventing)
• Save decisions to project_memories during planning
• Reference patterns from previous work
• Build knowledge that improves over time

This is the complete AI development system!

---

🎯 Skills Library

Browse, install, and manage Claude Code skills with ease!

What Are Skills?

Skills are reusable instruction sets that teach Claude specific capabilities. They can include:

• Coding patterns and best practices
• Tool usage workflows
• Domain-specific knowledge
• Development methodologies

Skill Types

Global Skills (~/.claude/skills/)

• Available in ALL projects
• Core skills: memory, initialize-project

Project Skills ({project}/.claude/skills/)

• Available only in that project
• Installed from templates or custom created

Community Skills (fetched from GitHub)

• Anthropic Official - 16 skills from anthropics/skills
• Superpowers - 20 skills from obra/superpowers

Using the Skills Command

# List all installed skills
/claude-os-skills

# Browse local templates
/claude-os-skills templates

# Install a template to your project
/claude-os-skills install rails-backend

# Create a custom skill
/claude-os-skills create

# View skill details
/claude-os-skills view <name>

# Delete a project skill
/claude-os-skills delete <name>

Community Skills (via Web UI)

1. Open the web UI at http://localhost:5173
2. Select your project
3. Click the Skills tab
4. Click Install Template
5. Switch to Community Skills tab
6. Browse skills from Anthropic Official and Superpowers
7. Click Install on any skill

Featured Community Skills

From Anthropic Official:

• pdf - Create, edit, and analyze PDF documents
• xlsx - Spreadsheet manipulation with formulas
• frontend-design - Production-grade UI components
• mcp-builder - Create MCP servers
• doc-coauthoring - Collaborative documentation

From Superpowers:

• test-driven-development - TDD workflow
• systematic-debugging - Four-phase debugging framework
• code-review - Rigorous code review process
• git-worktrees - Isolated development branches
• brainstorming - Structured ideation process

---

🎯 Spec Tracking & Kanban Board

NEW: Real-time auto-syncing Kanban board for Agent-OS specs!

Visual Kanban board showing specs, tasks, and progress tracking

What It Does

When you use Agent-OS to create specs with /create-spec, Claude OS automatically:

• 📋 Parses tasks.md files - Extracts all tasks, phases, dependencies, and metadata
• 🗄️ Stores in database - Tracks progress, completion, and time estimates
• 📊 Displays as Kanban - Visual board showing specs and tasks by status
• ⚡ Real-time sync - NEW! Auto-detects file changes and updates within 3 seconds
• 👀 File watching - Monitors agent-os/specs/ folder for changes
• ✅ Auto-refresh - Board polls every 3 seconds for live updates
• 🗃️ Archives completed specs - Keep your board focused on active work

Features

Real-Time File Watching (NEW!):

• Automatically monitors your agent-os/specs/ folder
• Detects changes to tasks.md and spec.md files
• 2-second debounce to batch rapid edits
• Auto-syncs to database within 3 seconds
• Frontend auto-refreshes every 3 seconds
• Total latency: ~6 seconds from file save to board update

Automatic Syncing:

• Syncs all specs from your project's agent-os/specs/ folder
• Tracks task metadata (estimated time, dependencies, risk level)
• Auto-detects completed tasks (marked with ✅ or [x] in tasks.md)
• Supports both checkbox format and classic format

Progress Tracking:

• Status auto-updates based on completion:
  • planning - No tasks completed yet
  • in_progress - Some tasks completed
  • completed - All tasks done
• Progress percentage calculated automatically
• Time estimates tracked (estimated vs actual minutes)

Archive Feature:

• Archive completed specs to keep your board clean
• Archived specs hidden by default but can be viewed
• Preserves all task history for future reference

API Endpoints

All spec tracking functionality is exposed via REST API:

# Get all specs for a project
GET /api/projects/{project_id}/specs

# Get all tasks for a spec
GET /api/specs/{spec_id}/tasks

# Update task status
PATCH /api/tasks/{task_id}/status
{
  "status": "in_progress",  # todo, in_progress, done, blocked
  "actual_minutes": 15
}

# Sync specs from agent-os folder (manual)
POST /api/projects/{project_id}/specs/sync

# Get Kanban board view
GET /api/projects/{project_id}/kanban?include_archived=false

# Archive/unarchive specs
POST /api/specs/{spec_id}/archive
POST /api/specs/{spec_id}/unarchive

# NEW: Real-time spec watcher control
GET /api/spec-watcher/status
POST /api/spec-watcher/start/{project_id}
POST /api/spec-watcher/stop/{project_id}
POST /api/spec-watcher/start-all

See: docs/guides/REALTIME_KANBAN_GUIDE.md for complete documentation.

How It Works

1. You create a spec with Agent-OS:
   /create-spec → agent-os/specs/2025-01-15-user-auth/

2. Spec Watcher detects the new folder:
   - Auto-starts when MCP server boots
   - Monitors agent-os/specs/ directory
   - 2-second debounce for batch changes

3. Auto-sync to database:
   - Reads tasks.md
   - Parses checkbox format: - [x] Task title
   - Extracts metadata, tasks, phases
   - Stores in SQLite database
   - ✅ Completes within 3 seconds

4. View in Kanban board (auto-refreshes every 3 seconds):
   - Todo: PHASE1-TASK1, PHASE1-TASK2
   - In Progress: PHASE2-TASK1
   - Done: PHASE1-TASK3, PHASE1-TASK4

5. As you work, agent-os updates tasks.md:
   - File watcher detects change
   - Auto-syncs to database
   - Board refreshes automatically
   - Total latency: ~6 seconds

6. Archive when complete:
   - Mark spec as archived
   - Keeps history but cleans up board

Database Schema

Two new tables track specs and tasks:

specs table:

• Stores spec metadata (name, path, status)
• Tracks total/completed tasks
• Calculates progress percentage
• Archive flag to hide completed specs

spec_tasks table:

• Individual tasks with codes (PHASE1-TASK1)
• Status (todo/in_progress/done/blocked)
• Time tracking (estimated vs actual)
• Dependencies between tasks
• Risk levels and phases

Example Usage

# Sync all specs for your project
curl -X POST http://localhost:8051/api/projects/1/specs/sync

# Response:
{
  "synced": 3,
  "updated": 0,
  "total": 3,
  "errors": []
}

# Get Kanban view
curl http://localhost:8051/api/projects/1/kanban

# Response shows:
# - Your specs with tasks
# - Tasks grouped by status
# - Progress percentages
# - Time estimates

This is the complete AI development system!

---

📂 Template System

claude-os/
├── templates/              # Shared templates
│   ├── commands/          # Slash commands (symlinked to ~/.claude/)
│   │   ├── claude-os-init.md
│   │   ├── claude-os-search.md
│   │   ├── claude-os-skills.md
│   │   ├── claude-os-lifecycle.md  # KB lifecycle management
│   │   └── ...
│   ├── skills/            # Global skills (symlinked to ~/.claude/)
│   │   ├── initialize-project/
│   │   └── memory/
│   ├── skill-library/     # NEW: Local skill templates
│   │   ├── general/       # General purpose skills
│   │   ├── rails/         # Ruby on Rails skills
│   │   ├── react/         # React/TypeScript skills
│   │   └── testing/       # Testing frameworks
│   └── project-files/     # Files created during /claude-os-init
│       ├── CLAUDE.md.template
│       └── .claude-os/
│           ├── config.json.template
│           └── hooks.json.template
├── cli/                   # CLI tools
│   └── claude-os-consolidate.sh
├── install.sh             # Quick setup script
└── start.sh               # Start services

Benefits:

• ✅ Update once, all projects benefit
• ✅ Symlinks mean instant updates
• ✅ Consistent across projects

---

📚 Managing Knowledge Bases

Via Web UI

1. Visit http://localhost:5173
2. Create Knowledge Base:
   • Click "Create Knowledge Base"
   • Choose type (Generic, Code, Documentation, Agent_OS)
3. Upload Documents:
   • Select KB from dropdown
   • Drag & drop files or click upload
   • Supports .md, .txt, .pdf, .py, .js, .ts, .json, .yaml
4. Query:
   • Type question in search box
   • View answer with source citations

Via CLI

# Search your project memories
/claude-os-search "how did we implement authentication?"

# Save a quick insight
/claude-os-remember "Fixed bug in user controller by adding validation"

# Full-featured save
/claude-os-save "Authentication Pattern" my-app-project_profile Architecture

Auto-Created KBs

When you run /claude-os-init, you get 4 knowledge bases:

1. {project}-project_memories

   • Claude's memory for decisions, patterns, solutions
   • Automatically saved during sessions
   • Searched at session start

2. {project}-project_profile

   • Architecture, coding standards, practices
   • Generated by initialize-project skill
   • Updated as project evolves

3. {project}-project_index

   • Automated codebase index
   • Tracks file structure
   • Updates on git commits (with hooks)

4. {project}-knowledge_docs

   • Your documentation
   • Auto-ingested during init
   • Add more via UI or CLI

---

🔬 Knowledge Lifecycle Management

As your knowledge bases grow, they can accumulate duplicates, outdated content, and fragmented information. The Knowledge Lifecycle Engine keeps your KBs healthy and focused.

Commands

/claude-os-lifecycle health [kb_name]         # Health report with recommendations
/claude-os-lifecycle dedup [kb_name]          # Scan and merge duplicate documents
/claude-os-lifecycle consolidate [kb_name]    # LLM-powered document merging
/claude-os-lifecycle archive [kb_name]        # Find stale docs, archive/restore
/claude-os-lifecycle logs [kb_name]           # View operation history

Deduplication

Scans all document embeddings to find near-duplicates using cosine similarity:

# Scan for duplicates (default threshold: 0.85)
/claude-os-lifecycle dedup my-project-project_memories

# Results show clusters of similar docs with options:
# [m] Merge - keep best, delete rest
# [c] Consolidate - LLM-merge into one comprehensive doc
# [s] Skip

Consolidation

Uses LLM to intelligently merge multiple related documents into a single comprehensive document:

/claude-os-lifecycle consolidate my-project-project_memories
# Previews source docs, then generates a merged version
# Preserves all unique info, eliminates redundancy
# Stores provenance metadata (consolidated_from)

Archival

Soft-archive stale documents without permanent deletion:

# Find documents older than 90 days
/claude-os-lifecycle archive my-project-project_memories

# Archived docs are excluded from search but can be restored
# Full restore available at any time

Health Reports

Get a comprehensive overview of your KB's health:

• Embedding coverage - How many docs have vectors
• Age distribution - Document freshness breakdown
• Top similar pairs - Preview of potential duplicates
• Recommendations - Actionable suggestions (dedup, re-index, archive)
• Growth timeline - Track KB growth over time

MCP Tools

Tool	Description
mcp__code-forge__kb_lifecycle_health	Health report with recommendations
mcp__code-forge__kb_lifecycle_dedup	Scan/merge duplicates
mcp__code-forge__kb_lifecycle_consolidate	LLM-powered document merging
mcp__code-forge__kb_lifecycle_archive	Archive, restore, list, find stale

API Endpoints

All under /api/kb/{kb_name}/lifecycle/:

Method	Path	Description
POST	/dedup-scan	Scan for duplicates (background for >500 docs)
POST	/dedup-merge	Merge duplicate documents
POST	/consolidate	LLM-powered consolidation (background)
GET	/health	Comprehensive health report
GET	/growth	Document growth timeline
POST	/archive	Archive documents
POST	/restore	Restore archived documents
GET	/archived	List archived documents
GET	/stale	Find stale documents
GET	/logs	Operation audit log

---

⚙️ Configuration

Environment Variables

# Provider (local = Ollama, openai = OpenAI API)
CLAUDE_OS_PROVIDER=local            # Default: local

# SQLite Database
SQLITE_DB_PATH=data/claude-os.db    # Default: data/claude-os.db

# Ollama (for local provider)
OLLAMA_HOST=http://localhost:11434  # Default: localhost:11434
OLLAMA_MODEL=llama3.2:3b            # Default: llama3.2:3b (lite model)
OLLAMA_EMBED_MODEL=nomic-embed-text # Default: nomic-embed-text

# OpenAI (for openai provider)
OPENAI_API_KEY=sk-...               # Required if using OpenAI
OPENAI_LLM_MODEL=gpt-4o-mini        # Default: gpt-4o-mini
OPENAI_EMBED_MODEL=text-embedding-3-small  # Default

# MCP Server
MCP_SERVER_HOST=0.0.0.0             # Default: 0.0.0.0
MCP_SERVER_PORT=8051                # Default: 8051

Project Configuration

Each project has .claude-os/config.json:

{
  "project_name": "my-app",
  "claude_os_url": "http://localhost:8051",
  "knowledge_bases": {
    "memories": "my-app-project_memories",
    "profile": "my-app-project_profile",
    "index": "my-app-project_index",
    "docs": "my-app-knowledge_docs"
  },
  "docs_settings": {
    "watch_paths": ["./docs", "./knowledge_docs"],
    "auto_ingest_patterns": ["*.md", "*.txt", "*.pdf"]
  },
  "tech_stack": "Ruby on Rails",
  "database": "MySQL"
}

---

📊 Performance

Native Ollama Setup:

• Response time: ~40 seconds per query
• GPU acceleration: Full Metal GPU on Apple Silicon
• Memory usage: 8-10GB (models + context)
• CPU usage: 12 cores (M4 Pro)

Why it's fast:

• Direct GPU acceleration (no virtualization)
• Efficient vector search in SQLite
• Optimized RAG engine with caching
• Single-file database with minimal overhead

---

🛠️ Scripts Guide

Installation & Setup

./setup-claude-os.sh - Unified Installer (Recommended)

./setup-claude-os.sh           # Interactive installation
./setup-claude-os.sh --demo    # Try the beautiful UI (no changes)
./setup-claude-os.sh --dry-run # Preview what would happen
./setup-claude-os.sh --help    # Show all options
./setup-claude-os.sh --version # Show version (v2.2.0)

Features:

• 🎨 Beautiful interactive UI with Charm CLI (gum) support
• 🛡️ Safety features: --demo, --dry-run, automatic config backups
• ☁️ Provider choice: Local (Ollama) or Cloud (OpenAI)
• 💨 Model choice: Lite (llama3.2:3b, 2GB) or Full (llama3.1:8b, 4.7GB)
• 🐧 Cross-platform: macOS and Linux (Ubuntu, Debian, Fedora, RHEL, Arch)

What it installs:

• ✅ Python virtual environment + dependencies
• ✅ Ollama + AI models (if local provider)
• ✅ Redis for caching
• ✅ MCP server configuration
• ✅ Commands and skills symlinked to ~/.claude/

Legacy Scripts

The old scripts redirect to the unified installer:

./install.sh  # → redirects to setup-claude-os.sh
./setup.sh    # → redirects to setup-claude-os.sh

Service Management

./start.sh or ./start_all_services.sh - Start Everything

./start.sh

Starts:

• 🔌 MCP Server (port 8051)
• 🎨 React Frontend (port 5173)
• 🤖 RQ Workers
• 💾 Redis
• 🧠 Ollama

./stop_all_services.sh - Stop All

./stop_all_services.sh

./restart_services.sh - Restart

./restart_services.sh

---

🗑️ Uninstalling Claude OS

To completely remove Claude OS from your system:

cd /path/to/claude-os
./uninstall.sh

The uninstall script removes:

• Command symlinks from ~/.claude/commands/
• Skill symlinks from ~/.claude/skills/
• MCP server config from ~/.claude/mcp-servers/
• Python virtual environment (venv/)
• Config files and logs
• Optionally: your knowledge base data

What it does NOT remove:

• The claude-os/ directory itself (delete manually with rm -rf)
• Ollama (see Ollama uninstall docs)
• Redis (brew uninstall redis on macOS)

Manual Uninstall:

If you prefer to uninstall manually:

# Remove symlinks
rm ~/.claude/commands/claude-os-*.md
rm -rf ~/.claude/skills/initialize-project
rm -rf ~/.claude/skills/memory
rm ~/.claude/mcp-servers/code-forge.json

# Remove Claude OS directory
rm -rf /path/to/claude-os

---

🐛 Troubleshooting

"Command not found: /claude-os-init"

Symlinks weren't created. Re-run:

cd /path/to/claude-os
./install.sh

"Connection refused to localhost:8051"

Claude OS server isn't running:

cd /path/to/claude-os
./start.sh

"Project already exists"

Project name is taken. Choose a different name or delete via UI at http://localhost:5173

Port Already in Use

# Find process on port 8051
lsof -i :8051

# Kill if needed
kill -9 <PID>

Ollama Issues

# Check if running
ollama list

# Start manually
ollama serve

# Check for model
ollama list | grep llama3.1

---

📁 Project Structure

claude-os/
├── templates/              # Shared templates system
│   ├── commands/          # Slash commands
│   ├── skills/            # Global skills
│   ├── skill-library/     # Local skill templates (NEW)
│   └── project-files/     # Files created during init
├── cli/                   # CLI tools
│   └── claude-os-consolidate.sh
├── app/                    # Backend application
│   ├── core/              # Core modules
│   │   ├── sqlite_manager.py
│   │   ├── rag_engine.py
│   │   ├── skill_manager.py    # Skills management
│   │   ├── session_parser.py   # Session parsing
│   │   ├── insight_extractor.py # Insight extraction
│   │   ├── knowledge_lifecycle.py # KB lifecycle engine (dedup, archive, etc.)
│   │   └── ...
│   └── db/                # Database schemas
├── frontend/              # React UI (Vite)
│   ├── src/
│   │   ├── components/
│   │   │   ├── SkillsManagement.tsx  # NEW: Skills UI
│   │   │   └── ...
│   │   └── pages/
│   └── public/
│       └── assets/
├── mcp_server/           # MCP Server (HTTP)
│   └── server.py         # FastAPI + MCP endpoints
├── data/                 # SQLite database
│   └── claude-os.db
├── logs/                 # Service logs
├── install.sh            # Quick setup script
├── start.sh              # Start services
└── README.md             # This file

---

📖 Additional Documentation

Getting Started

• templates/README.md - 📂 Template system documentation

Core Features

• docs/guides/REALTIME_KANBAN_GUIDE.md - ⚡ NEW! Real-time Kanban board (auto-sync, file watching, API reference)
• docs/SELF_LEARNING_SYSTEM.md - 🧠 How Claude learns automatically
• docs/REAL_TIME_LEARNING_GUIDE.md - Real-time learning usage

Technical Documentation

• docs/API_REFERENCE.md - 🔌 Complete API Reference (all endpoints, examples, authentication)
• docs/HYBRID_INDEXING_DESIGN.md - ⚡ Hybrid indexing architecture
• README_NATIVE_SETUP.md - Detailed native setup
• NATIVE_VS_DOCKER_DECISION.md - Why native Ollama
• PERFORMANCE_TEST_RESULTS.md - Benchmark results

---

🤝 Contributing

Claude OS is open source. Feel free to:

• Modify for your specific needs
• Add new commands and skills
• Optimize RAG strategies
• Contribute improvements back

---

🙏 Acknowledgments

Agent-OS Integration

Claude OS optionally integrates with Agent-OS by Builder Methods (CasJam Media LLC).

• Project: Agent-OS - Spec-driven development workflow system
• Author: Builder Methods (CasJam Media LLC)
• License: MIT
• Repository: https://github.com/builder-methods/agent-os

Agent-OS provides 8 specialized agents for structured feature planning and implementation. We're grateful to Builder Methods for creating such powerful tools and for licensing them under MIT, making this integration possible.

If you find Agent-OS valuable, consider:

• ⭐ Starring their repository
• 📣 Sharing it with other developers
• 🤝 Contributing to their project

---

📄 License

MIT License - Use it freely!

Note: This project (Claude OS) is MIT licensed. Agent-OS, when installed, is a separate project also MIT licensed by Builder Methods (CasJam Media LLC). See the Agent-OS repository for their specific license terms.

---

Claude Code + Claude OS = Invincible! 🚀
Built by AI coders, for AI coders