An open-source AI agent that runs 24/7 on your own server, uses any LLM as its "brain", remembers across sessions and writes its own skills from the work it has done. This guide takes you from install to real-world usage, situation by situation.
Unlike agents that just "connect to everything", Hermes bets on learning from experience — the longer it runs, the better it gets. Launched 25/02/2026 and overtook OpenClaw on 10/05 to lead global inference volume on OpenRouter.
After every task that uses 5+ tool calls, the agent runs a "reflection" step and produces a reusable SKILL.md skill file. With 20+ skills, similar tasks later run ~40% faster (in tokens & time).
A background process that runs weekly, scoring, rewriting and pruning low-performing skills — keeping the skill library lean and "alive".
Stored in SQLite with FTS5 full-text search, remembering across sessions. The agent "distills" its own memory, plus Honcho builds an ever-deeper model of the user.
One agent, one memory, every surface: Telegram, Discord, Slack, WhatsApp, Signal, Email, Teams… plus CLI/TUI, Web UI and IDE integrations (VS Code, Zed, JetBrains via ACP).
Hand work off to independent sub-agents (their own context + terminal), running in parallel. Programmatic Tool Calling via execute_code folds many steps into a single call.
Schedule in plain speech: "every 9am, summarize the inbox and send it to Slack" — runs in the background via the gateway, delivering results to any platform.
Self-hosted, MIT license, no telemetry, your data stays on your machine. Can run local models via LM Studio — sensitive data never leaves the box.
Six terminal backends: local, Docker, SSH, Singularity, Modal, Daytona — with container hardening and namespace isolation to run commands safely.
Skills follow the open agentskills.io standard, shared via the Skills Hub. MCP support. Research-ready: batching, trajectory export, RL with Atropos.
Two common paths from two tutorial videos: NetworkChuck builds it manually on a VPS to understand it deeply, while Charlie Chang uses a 1-click deploy for speed.
Pick the KVM 2 plan, Ubuntu OS. This is the server where Hermes will run continuously 24/7.
Paste the SSH command from the Hostinger dashboard into your terminal, then enter the root password to log in.
Copy the one-line command from the Hermes site and paste it — the script installs everything for you.
When hermes setup runs, pick a provider: Nous Portal, OpenAI, Grok, OpenRouter, or a local model (LM Studio + Qwen).
Create a new bot (its name must end in bot), copy the token and paste it into the terminal.
Use the @userinfobot bot to get your Telegram User ID, then add it to the allowlist — so strangers can't control your agent.
Choose "system service" so the agent starts automatically when the server restarts. On a personal VPS you can run it as the root user.
Use the Hostinger link → Deploy → checkout (~$8/month). Hostinger provisions the VPS + Docker container for Hermes automatically.
Fill these in the dashboard — almost no terminal typing needed:
Open the Terminal in the dashboard and run:
Open the Telegram bot → type /start → copy the pairing code → paste it into the terminal. Done — the agent is ready to take commands from your phone.
Three systems run alongside the main conversation loop to create the "an agent that grows with you" experience.
Info about you: name, preferences, team, work context.
Environment & technical info: IPs, config, system notes.
The agent's persona / identity — it decides its "personality" in conversation.
All three files are loaded into the system prompt automatically at the start of each session. When a file fills up, Hermes "curates" it to prioritize the important information. Every 10 turns, a background task updates the memory.
Send a character description into the chat and the agent writes it into SOUL.md, holding the persona throughout:
Give it instructions + an API link, and the agent learns how to call it and saves it as a skill:
Send commands to the agent via Telegram and the terminal at the same time, and it handles them in parallel. A message queue mechanism finishes the current task first instead of getting interrupted.
Attach any MCP server (via URL or a local command) to extend its capabilities. A Nous-vetted catalog allows 1-click installs and tool filtering for safe use.
Click each card to see the configuration and a sample prompt. Tip: for each purpose, create a separate "profile" so the memory doesn't get mixed up.
fal_key, then let the agent create a skill that calls the fal.ai API (CogVideoX) to generate video from an image.delegate_task (up to 3 children at once by default), and the parent posts cards to a built-in Kanban board for the children to pick up.Both use the open SKILL.md skill standard, but the philosophies differ: OpenClaw connects to as much as possible, Hermes learns from experience.
| Criteria | Hermes Agent | OpenClaw |
|---|---|---|
| Philosophy | Learns from experience, writes its own skills | Connects to the most platforms |
| Skills | Self-built from interaction + Curator cleanup | Downloads ready-made skills from a marketplace |
| Memory | Self-curated hard limit, doesn't balloon | Prone to bloat over time |
| Stability | "Like a product" | Has had a security crisis before |
| Team | Nous Research — an AI research lab | Tooling-focused |
| Security | A "less is more" philosophy | Larger attack surface |