OptimusHQ

A self-hosted, multi-tenant platform for running multiple Claude Code agents across projects simultaneously. Agents can delegate tasks to each other and communicate across sessions. Manage agents, projects, skills, and sessions from a single dashboard with real-time streaming, browser automation, and a full kanban workflow. Each user gets their own isolated workspace with cloned agents, skills, and settings.

Built with React, Express, WebSocket, and SQLite. Runs on any machine with Node.js and the Claude CLI.

Why

Claude Code is powerful but limited to one terminal session at a time. This project turns it into a multi-project, multi-agent platform where you can run concurrent agents across different codebases, each with their own context, memory, skills, and tools.

Features

Multi-Project Workspace

• Create projects with auto-generated folders at ~/projects/<slug>/ or link existing codebases
• Clone git repositories directly from the UI or via MCP tool
• Each project gets its own dev server port (3100-3999), preview URL, git settings, and server config
• Projects are fully isolated -- agents work within their assigned project directory

Agents

• 5 built-in agent personas: Builder, Researcher, Debugger, Writer, DevOps
• Create custom agents with tailored system prompts
• Per-agent model selection: Haiku, Sonnet, Opus
• Switch agents mid-session or assign different agents to different tasks
• Each session spawns its own Claude CLI process -- agents work concurrently in the background

Chat

• Real-time streaming over WebSocket with structured NDJSON output
• Tool use is displayed inline (Bash commands, file edits, browser actions)
• Per-message model override (switch between Haiku/Sonnet/Opus on the fly)
• Message queue with pending message handling across page navigation
• Interrupt running agents at any time
• Session-based isolation with WebSocket message queuing for reliability

Mission Control (Kanban Board)

• 4-column board: Backlog, In Progress, Review, Done
• Drag-and-drop session cards between columns
• Filter by project with color-coded cards
• Activity feed with timestamped status changes and actor attribution

Permission Modes

Each session can run in one of three modes:

• Execute (default): Full access to all tools -- file writes, bash, browser, everything
• Explore (read-only): Can only read files, search, and browse -- all write tools are blocked
• Ask (confirmation required): Agent must describe planned changes and wait for user approval before executing

Skills

• Modular prompt instructions that can be toggled per session
• 5 built-in skills: Code Review, Concise Output, Testing, WordPress Plugin Dev, Email Marketing
• Create custom skills with full prompt definitions
• Import skills from skills.sh or any GitHub URL
• Chat-based skill creation: Describe skills in natural language or paste a GitHub repo URL -- the agent imports them conversationally using the create_skill MCP tool
• Scope skills globally or per project
• File glob patterns to target specific file types

Security Warning: Always review skill instructions before installing. Malicious skills can contain instructions that exfiltrate sensitive data (like .env files, API keys, or credentials) to external servers. Only install skills from trusted sources and inspect the prompt content carefully.

APIs

• Register external APIs with base URL, auth config, and endpoint documentation
• Auth types: Bearer token, custom header, query parameter, HTTP Basic
• Generate API configurations from a text description or documentation URL (uses Claude)
• Enable/disable APIs per session -- credentials and docs are injected into the agent's context

MCP Servers (Model Context Protocol)

Two built-in MCP servers:

Chrome DevTools -- Full browser automation:

• Navigate pages, click elements, fill forms, take screenshots
• Read console logs, inspect network requests
• Emulate devices, throttle network, change viewport
• Performance tracing and Core Web Vitals

Project Manager -- Project and memory management:

• create_project / clone_project / list_projects
• get_project_memory / update_project_memory
• get_server_config / update_server_config
• create_skill -- Create skills programmatically from agent conversations
• create_mcp -- Create MCP servers programmatically with encrypted env vars
• delegate_task / send_message -- Cross-session agent delegation
• add_memory_entry / search_memory / list_memory_entries -- Structured memory management
• get_project_status -- Check session activity and recent messages
• scrape_url -- Fetch a URL and extract clean content (title, text, metadata)
• crawl_links -- Extract and filter links from a page for article/URL discovery

Internal vs User MCP servers: System MCP servers (like Project Manager) are hidden from the UI to keep the interface clean. Users can add custom MCP servers with command, args, and environment variables.

Chat-based MCP import: Describe an MCP or paste a GitHub/npm URL in the MCPs page -- the agent researches it and creates the MCP configuration using the create_mcp tool. Environment variables (API keys) are encrypted at rest using AES-256-CBC.

Security Warning: MCP servers execute code on your system. Only install MCPs from trusted sources. Review the source code before installation -- malicious MCPs can access files, environment variables, and network resources. Be especially careful with MCPs that request sensitive environment variables.

Memory

Three levels of persistent memory that work together to give agents long-term context:

Session Memory

• Auto-summarized every 5 messages using Haiku (cost-optimized)
• Pulls the last 20 messages and generates a sub-200-word summary
• Runs asynchronously after each agent response -- does not block the chat
• Includes pinned facts for critical info that should never be forgotten
• Can be disabled per project via the auto_summarize flag

Project Memory

• Shared across all sessions in a project -- persistent context that survives session boundaries
• Agents read and write it via get_project_memory / update_project_memory MCP tools
• Use it for architecture decisions, known issues, working notes -- anything future sessions should know
• Editable from the UI or programmatically by agents

Memory Entries (Structured)

• Discrete, searchable items with categories: decision, feature, bug, content, todo, context
• Auto-extracted every 10 messages -- Haiku reviews recent conversation and pulls up to 3 entries
• Deduplication: skips entries with duplicate titles within 24 hours
• Full-text search via SQLite FTS5 with support for phrases, AND/OR operators, and wildcards
• Agents can manually create entries via add_memory_entry and search with search_memory
• Tags support for additional organization and filtering

How memory flows into agent context:

1. On each session start, the system prompt is assembled with project memory, the last 5 memory entries, and the session summary
2. Agents are informed about available memory MCP tools (search_memory, add_memory_entry, list_memory_entries, get_project_memory, update_project_memory)
3. All summarization and extraction runs async using Haiku to keep costs low and avoid blocking the chat

Server Config

Dedicated per-project field for dev server management:

• Store startup commands, required services, health check commands, and recovery steps
• Agents automatically check if the dev server is running at session start
• If the server is down, agents follow the stored config to bring it back up
• Agents save server config via MCP after setting up new projects
• Editable in project settings UI

File Explorer

• Browse project directory tree with smart filtering (ignores node_modules, .git, dist, etc.)
• Read and edit files with syntax-highlighted CodeMirror editor
• Create and delete files and directories
• Path traversal protection -- stays within project boundaries

Cross-Agent Delegation

• Agents can delegate tasks to other agents via delegate_task MCP tool
• Send messages to existing sessions with send_message
• Spawns background sessions that work autonomously
• Track delegated work through the activity feed

Source Control

• Git status, staging, unstaging, commits, diffs -- all from the sidebar
• Branch switching and commit log viewer
• Push/pull with safety controls:
  • Per-project push disable toggle (pull-only mode)
  • Protected branches list (prevent pushes to main, production, etc.)
• Auto-detect git origin URL from existing repos

Live Preview

• Subdomain proxy: https://<project-name>.<base-domain>/ proxies to the project's dev port
• Static preview: /preview/<project-name>/ serves files directly from the project folder
• Preview URLs are shown in project settings and injected into agent context
• Base domain is configurable in Settings > Platform (admin only)

Settings

• Platform (admin): Base domain for project subdomain URLs (e.g., example.com makes projects accessible at <project>.example.com)
• Default model selection (Haiku, Sonnet, Opus)
• Default permission mode
• Tool allow/disallow lists (restrict which tools agents can use)
• Theme accent color (12 options)
• GitHub token for authenticated git operations

Multi-Tenancy

• Multiple users with complete data isolation
• Per-user agents, skills, projects, API keys, MCPs, and settings
• Admin user can switch between users via sidebar dropdown to view their data
• New users get cloned default agents, skills, and their own General project
• WebSocket authentication with user-scoped message delivery
• bcrypt password hashing with database-persisted auth tokens
• All 13 API routes scoped by user_id with cross-user access blocked

Project Isolation (Security)

• Agents are restricted to their assigned project directory via Claude Code hooks
• File operations (Read, Write, Edit) are validated against the project path
• Prevents cross-project file access even with --dangerously-skip-permissions
• /tmp is allowed for temporary files
• General project (no path) has unrestricted access
• Hook is auto-configured on server startup

WhatsApp Integration

• Chat with agents via WhatsApp on your phone
• Admin connects WhatsApp via QR code in Settings
• Users are identified by phone number (must be set in profile)
• Messages routed to user's account based on registered phone
• Uses Baileys (lightweight WebSocket client) -- no Puppeteer or browser needed

Mobile Responsive

• Collapsible sidebar with hamburger menu on mobile
• Responsive layouts across all pages
• Touch-friendly UI elements
• Works on phones and tablets

Blog

• Built-in blog system with markdown support
• Migrated from Sanity CMS
• SEO-friendly URLs

Database Backups

• Daily automated SQLite backups
• 7-day retention policy
• Backup script included (backup-db.sh)

Architecture

optimushq/
├── client/             # React 18 + Vite + Tailwind CSS
│   ├── src/pages/      # 10 pages (Chat, Board, Agents, Skills, APIs, MCPs, Settings, etc.)
│   ├── src/components/ # Organized by feature
│   └── src/hooks/      # 13 custom React hooks
├── server/             # Express + WebSocket + SQLite
│   ├── src/routes/     # REST API (projects, sessions, skills, APIs, git, files, etc.)
│   ├── src/claude/     # CLI spawning and context assembly
│   ├── src/tools/      # MCP server implementations
│   └── src/ws/         # WebSocket message handling
├── shared/             # TypeScript interfaces shared between client and server
├── mcp-config.json     # Generated MCP configuration (dynamic)
├── ecosystem.config.cjs # PM2 config
└── start-chrome.sh     # Headless Chrome launcher

How a message flows:

1. User sends message in chat UI
2. Server saves message to SQLite, assembles full context (agent prompt + project info + skills + APIs + memory + server config + permission rules)
3. Spawns claude CLI with --output-format stream-json and --mcp-config
4. Streams NDJSON response chunks back over WebSocket in real-time
5. Tool use (file edits, bash commands, browser actions) displayed inline
6. Response saved to DB, auto-summarization triggers every 5 messages

Prerequisites

• Node.js 20+
• Claude CLI installed and authenticated (claude --version should work)
• Chrome/Chromium (optional, for browser automation via DevTools MCP)

Quick Start

Option 1: npx (fastest)

npx @goranefbl/optimushq

Opens at http://localhost:3001. Login with admin / admin.

Important: Do not run as root. Claude CLI blocks --dangerously-skip-permissions for root users. Create a non-root user:

sudo useradd -m claude
sudo su - claude
npx @goranefbl/optimushq

To enable WhatsApp, go to Settings and scan the QR code with WhatsApp.

Option 2: Clone and run

git clone https://github.com/goranefbl/optimushq.git
cd optimushq

npm install
npm run dev

Opens at http://localhost:5173 (Vite proxies API calls to port 3001).

Default login: Set via AUTH_USER and AUTH_PASS environment variables (defaults to admin / admin)

Production

npm run build

# Option 1: Run directly
AUTH_USER=admin AUTH_PASS=changeme PORT=3001 npx tsx server/src/index.ts

# Option 2: PM2 (recommended)
pm2 start ecosystem.config.cjs

The app runs at http://localhost:3001 with the built frontend served statically.

Environment Variables

Variable	Description	Default
PORT	Server port	3001
AUTH_USER	Login username	admin
AUTH_PASS	Login password	admin
BASE_DOMAIN	Fallback domain for project subdomains (can also be set in Settings)	localhost

Default Seed Data

On first run the server creates:

• 5 agents: Builder (default), Researcher, Debugger, Writer, DevOps
• 5 skills: Code Review, Concise Output, Testing, WordPress Plugin Dev, Email Marketing
• 2 MCP servers: Chrome DevTools (user-visible), Project Manager (internal)
• 1 project: General (for unscoped chats)

Database

SQLite with WAL mode. 19 tables covering users, auth tokens, projects, agents, sessions, messages, skills, APIs, MCP servers, memory entries (with FTS5 full-text search), activity log, blog posts, and settings. Schema is auto-created on first run with automatic migrations for new columns.

Automated Backups

A backup script is included at scripts/backup-db.sh that:

• Creates safe backups using SQLite's backup API (works while server is running)
• Compresses backups with gzip
• Retains 7 days of backups automatically

To enable daily backups, add to crontab (crontab -e):

0 3 * * * /path/to/optimushq/scripts/backup-db.sh >> /home/claude/backups/backup.log 2>&1

Backups are stored in /home/claude/backups/ (outside the repo for security).

Optional: HTTPS with Cloudflare Tunnel

cloudflared tunnel --url http://localhost:3001

Optional: Subdomain Proxy with Cloudflare

To enable project subdomains (e.g., myproject.yourdomain.com), set up Cloudflare DNS:

1. Add a wildcard DNS record in Cloudflare:

   • Type: A
   • Name: *
   • Content: <your-server-ip>
   • Proxy status: Proxied (orange cloud)

2. Add the root domain (if not already):

   • Type: A
   • Name: @
   • Content: <your-server-ip>
   • Proxy status: Proxied

3. Set the base domain in OptimusHQ:

   • Go to Settings > Platform
   • Enter your domain (e.g., yourdomain.com)
   • Save

4. SSL/TLS settings in Cloudflare:

   • Go to SSL/TLS > Overview
   • Set encryption mode to "Full" or "Full (strict)"

Projects will now be accessible at https://<project-name>.yourdomain.com/, automatically proxied to each project's dev port.

Optional: Chrome for DevTools MCP

./start-chrome.sh

Launches headless Chrome with remote debugging on port 9222. Enables browser automation tools for all agents.

Contributing

Contributions welcome. The codebase is TypeScript end-to-end with shared types between client and server. Key areas:

• server/src/claude/context.ts -- How agent context is assembled
• server/src/tools/project-manager-mcp.ts -- MCP tool implementations
• client/src/pages/ -- UI pages
• server/src/routes/ -- API endpoints

License

MIT