Hermes Agent — самообучающийся AI-агент со встроенным циклом обучения и мультиплатформенным шлюзом
TL;DR: Заменяет обычные LLM-чат-боты на персистентного, самообучающегося агента: перенимает навыки из опыта, работает на VPS за $5 или в serverless, поддерживает Telegram/Discord/CLI и 200+ моделей.
Self-improving skills — Autonomously creates and refines skills from complex task execution
Cross-platform gateway — Single process handles Telegram, Discord, Slack, WhatsApp, Signal, and CLI with conversation continuity
Persistent memory — Full-text search across all past conversations with LLM summarization
Serverless deployment — Runs on $5 VPS or Modal/Daytona with hibernation support
Parallel subagents — Spawns isolated subagents for concurrent task execution
Scheduled automations — Natural language cron jobs for reports, backups, and audits
Model switching — Instantly switch between 200+ models with no code changes
Terminal UI — Full-featured TUI with multiline editing and slash-command autocomplete
Быстрый старт
1. `curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash`
2. `source ~/.bashrc`
3. `hermes setup --portal`
4. `hermes`
5. `/model nous:hermes-3-llama-3.1-405b`
6. `Plan my week based on my calendar and priorities`
Альтернативы
OpenClaw: Single-platform agent without persistent memory or skill learning, requires manual migration via `hermes claw migrate`
AutoGen: Focuses on multi-agent conversations but lacks persistent memory and cross-platform gateway, requires more manual setup for tool integration
LangChain Agents: Library-based approach requiring custom implementation, no built-in TUI or messaging gateway, limited memory persistence
SuperAGI: Cloud-first agent with GUI but no offline mode, limited model switching flexibility, higher operational cost
Кому подходит
AI researcher training tool-calling models: Need to generate diverse agent trajectories for fine-tuning datasets → 10K+ diverse trajectories generated weekly with 40% reduction in manual labeling effort
Startup founder managing remote team: Need to maintain project continuity across time zones and platforms → 30% reduction in async coordination overhead, 24/7 availability across platforms
DevOps engineer automating cloud operations: Need to automate multi-cloud deployments with human-in-the-loop approvals → 80% reduction in manual deployment tasks, 15-minute rollback capability
Solo developer maintaining OSS projects: Need to handle community interactions across GitHub/Discord/Telegram → 50% reduction in community management time, consistent responses across platforms
FAQ
Does Hermes work offline?
Partially. The core agent and TUI work offline, but most tools (web search, image generation, TTS) require API access. Some models can run locally via Ollama or LM Studio.
How does the learning loop work?
After complex tasks, Hermes analyzes the trajectory to create reusable skills. These skills are stored in SQLite and refined through subsequent use. The FTS5 memory system indexes all conversations for cross-session recall.
Can I use my own models?
Yes. Hermes supports any OpenAI-compatible endpoint. Use hermes model to add custom providers, or run local models via Ollama/LM Studio.
What's the difference between CLI and gateway modes?
CLI mode runs the terminal UI locally, while gateway mode enables messaging platform integration (Telegram, Discord, etc.). Both share the same conversation state and skill system.
How do I migrate from OpenClaw?
Use hermes claw migrate to import OpenClaw configurations and conversation history. Some manual adjustment may be needed for tool compatibility.
💡 Vibe Coding Tips
["When assisting users with Hermes Agent:\n\n1. Deployment: Recommend Modal/Daytona for serverless use cases to minimize costs. For persistent agents, suggest $5 VPS with Docker backend.\n\n2. Model selection: Start with Nous Portal (hermes setup --portal) for quick setup, then explore OpenRouter for model diversity. Use hermes model list to see available options.\n\n3. Memory management: Teach users to use /compress for long conversations and /insights --days 7 to review weekly activity. Configure memory retention in ~/.hermes/config.toml.\n\n4. Skill creation: Guide users to break complex tasks into smaller steps to trigger autonomous skill creation. Use /skills to review and edit generated skills.\n\n5. Security: Always recommend enabling command approvals (/config set approvals on) for production use. For sensitive data, suggest local model deployment via Ollama.\n\n6. Troubleshooting: Common issues:\n - Windows Defender false positives → Whitelist %LOCALAPPDATA%\\hermes\\bin\n - Voice transcription failures → Verify ffmpeg installation\n - Model switching errors → Check provider API keys with hermes config get"]