• deepagents Architecture — Compiled graph runtime with fine-grained state management
• Multi-Agent Types — Core, fast, and search agents
• Plugin System — @register_agent("id") decorator for custom agents
• Streaming Output — Native SSE support
• Sub-agents — Multi-level delegation
• Thinking Mode — Extended thinking for Anthropic models
• Scheduled Tasks — Cron, interval, date, and manual task triggers with persisted scheduler state
• Human-in-the-Loop — Approval system with countdown timer, auto-extension, and urgent-state styling
• Persona Presets — Reusable persona configuration with permissions and runtime binding

• Multi-Provider Models — OpenAI, Anthropic, Google Gemini, and Kimi
• Full CRUD — Create, edit, delete, reorder, and batch import models via UI
• Channel Routing — Route the same model through different channels with model_id
• Role-based Access — MODEL_ADMIN permission and per-role model visibility
• Cross-session Memory — Native MongoDB-backed memory system
• Dual Skills Storage — File system plus MongoDB backup
• GitHub Sync — Import custom skills from GitHub
• Skill Marketplace — Browse, install, publish, and manage skills in bulk

• System + User MCP — Global and per-user MCP configuration
• Encrypted Storage — API keys encrypted at rest
• Dynamic Tool Caching — Cache MCP tools with manual refresh
• Multiple Transports — SSE and HTTP
• Permission Control — Transport-level access policies
• Sandbox Integration — Daytona and E2B execution support
• Built-in Tools — File reveal, project reveal, upload URL, env vars, audio transcription, persona preset tools, and more

• File Library — Browse revealed files with code preview, favorites, and project-based filtering
• Rich Previews — PDF, Word, Excel, PPT, Markdown, Mermaid, Excalidraw, images, and video playback
• Project Folders — Organize sessions into projects with drag-and-drop
• Session Sharing — Generate public share links for conversations
• Feedback — Thumbs rating, text comments, session linking, and run-level stats
• Notifications — In-app notification storage and delivery hooks

• Realtime — Redis + MongoDB dual-write, WebSocket, auto-reconnect, and shared-session updates
• Task Runtime — Local or Redis-backed arq execution with embedded worker lifecycle
• Security — JWT, RBAC, bcrypt, OAuth (Google/GitHub/Apple), email verification, CAPTCHA, and sandbox controls
• Observability — LangSmith tracing, structured logging, health checks, and distributed memory diagnostics
• Channels — Native Feishu integration plus an extensible multi-channel architecture
• Frontend Stack — React 19, Vite 6, TailwindCSS 3.4, PWA workers, Capacitor mobile builds, Tauri desktop shell, rich rendering, and responsive layouts
• i18n — English, Chinese, Japanese, Korean, and Russian

Edit the .env file with the following recommended settings:

# Recommended: Set a stable JWT secret (auto-generated on each restart if unset, invalidating existing sessions)
JWT_SECRET_KEY=your-stable-secret-key

# Recommended: Set MCP encryption salt (auto-generated on each restart if unset, invalidating saved MCP configs)
MCP_ENCRYPTION_SALT=your-stable-encryption-salt

# Optional: Configure MongoDB connection
MONGODB_URL=mongodb://localhost:27017
MONGODB_DB=agent_state
MONGODB_USERNAME=admin
MONGODB_PASSWORD=your-mongo-password

# Optional: Configure Redis connection
REDIS_URL=redis://localhost:6379/0
REDIS_PASSWORD=your-redis-password

# Optional: enable scheduled tasks
ENABLE_SCHEDULED_TASK=true

# Optional: choose task execution backend
TASK_BACKEND=arq  # local or arq

::: tip LLM models are configured through the Model Config UI after deployment — no environment variables needed. :::