In 2026, Hermes Agent (Nous Research, roughly 170k GitHub stars) became one of the hottest open-source self-evolving AI agents — cross-session memory, auto-built Skills, and Gateway access to Telegram and 20+ platforms. Stuck on install errors, API keys, or a silent Telegram bot? This guide gives you one command for the full stack, a six-step runbook to hermes gateway 24/7, and a MacBook / VPS / Mac Mini M4 rental comparison for production hosting.
Hermes Agent is not a chat wrapper. Under the MIT license it is a always-on autonomous agent that generates reusable Skills after completing task types (community reports show roughly 38% token savings on repeat work). It persists context through a SESSION → MEMORY.md → USER.md three-layer memory model.
Compared with orchestration frameworks like LangGraph / CrewAI or channel gateways like OpenClaw, Hermes stands out for native Skill self-evolution, cross-session persistent memory, and a single curl install.
This article does not rehash why you might switch to Mac (see our VPS migration write-up) or memory architecture deep dive (see the three-layer memory hardware guide). It focuses on the install and deploy runbook: macOS / Linux / WSL2 one-command script, hermes doctor, LLM config, first conversation, Telegram Bot setup, and Gateway daemon.
Pain point: Docs are scattered — newcomers do not know whether to install dependencies or configure API keys first.
Pain point: After install, hermes: command not found — usually PATH not refreshed.
Pain point: Telegram works in DMs but not groups — Group Privacy and re-invite order trip people up.
Pain point: Closing a laptop lid kills Gateway and breaks Skill compounding.
Pain point: Headless VPS lacks native macOS browser automation paths (Camoufox and similar).
What this delivers: A six-step checklist from curl to verifiable hermes gateway start.
The official install.sh uses uv to pull Python 3.11, Node.js 22, ripgrep, ffmpeg, and clones the repo to ~/.hermes/hermes-agent/. The table below is a sizing reference (compiled from Nous Research install docs and community deployment notes, June 2026).
| Item | Minimum | Recommended (24/7 + Gateway) |
|---|---|---|
| OS | macOS 12+ / Ubuntu 20.04+ / WSL2 | macOS 14+ (Apple Silicon) or Ubuntu 24.04 LTS |
| RAM | 4 GB (Gateway + cloud API only) | 16 GB; 32 GB for local inference |
| Disk | ~1.5 GB (including Skills cache) | 20 GB SSD+ |
| Network | Access to GitHub raw and LLM APIs | Stable low-latency egress (matters for Telegram UX) |
"The hardest part of installing Hermes is often not the command itself — it is picking a machine that will not sleep or drop offline at 2 a.m."
# Official one-command install — macOS / Linux / WSL2 curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash # Headless servers can skip browser automation components # curl -fsSL .../install.sh | bash -s -- --skip-browser source ~/.zshrc # or source ~/.bashrc
Install and refresh PATH: After you see ✅ Hermes Agent installed at ~/.hermes, run source ~/.zshrc, open a new terminal, then run hermes --version.
Health check: hermes doctor validates Python, Node, ripgrep, ffmpeg, PATH, API keys, and about 12 other items. Fix anything it flags.
Configure LLM: New users can run hermes setup; or use hermes model to pick OpenRouter / Anthropic / OpenAI / Ollama, with keys written to ~/.hermes/.env (e.g. hermes config set OPENROUTER_API_KEY sk-or-xxxx).
First conversation: Run hermes in the CLI and try a task like "summarize today's AI news" or "count files in the current directory" to confirm the toolchain works.
Telegram Bot: Use /newbot in @BotFather for a token; use @userinfobot for your numeric ID; run hermes gateway setup and pick Telegram, or set TELEGRAM_BOT_TOKEN and TELEGRAM_ALLOWED_USERS.
24/7 daemon: Test with hermes gateway, then run hermes gateway install && hermes gateway start; macOS uses launchd, Linux can use sudo hermes gateway install --system + systemd.
By default, bots in groups only respond to / commands. To @mention the bot in group chat, disable Group Privacy in BotFather and remove and re-invite the bot — the most common "installed but silent" cause in production.
Common errors: API key not set → rerun hermes model; ModuleNotFoundError: dotenv → confirm you are using ~/.hermes/hermes-agent/venv/bin/hermes, not system Python; config drift after upgrade → hermes config check && hermes config migrate.
Hermes Skill compounding depends on a host that stays online long term. Use this table for a quick decision (availability figures are operational feel, not SLA guarantees):
| Platform | 24/7 availability | Skill compounding | Best for |
|---|---|---|---|
| Personal MacBook | Drops when lid closes (~60%) | Frequent interruptions | POC only |
| Raspberry Pi 4B 8GB | ~88% | I/O bottleneck, long tasks timeout | Hobby use |
| x86 VPS 4GB | ~99.5% | Stable but no macOS toolchain | Linux-only users |
| Mac Mini M4 rental | Colo-hosted, 30-day uptime | UMA + native launchd daemon | Telegram 24/7 + macOS Skill paths |
Tip: M4 unified memory lets Gateway, browser automation, and SQLite WAL share a high-bandwidth pool. Community monitoring shows Gateway at roughly 4 GB with headroom on 16 GB models. After renting remotely, SSH in and run the same install commands as on a local machine.
~/.hermes/hermes-agent/; global hermes command added to PATH by the install script.hermes update.~/.hermes/skills/; host downtime breaks compounding — directly tied to the 24/7 requirement.For weekend experiments, a laptop plus cloud API is enough. Once Hermes becomes a colleague you talk to daily, lid sleep, macOS gaps on VPS, and group privacy pitfalls compound together. Buying a Mac Mini works, but renting for 30 days to validate your workflow before committing is the lowest-cost decision — see rental rates for models.
Home Wi-Fi drops, oversold neighbor VPS, and Raspberry Pi disk I/O all kill Gateway at the wrong moment — Telegram users only see "the bot died again." For production setups that need Hermes running 24/7 with native macOS install and launchd, NodeMini Mac Mini M4 cloud rental is usually less painful than a "good enough laptop plus manual restarts" — you focus on agent compounding, not fixing hermes gateway status at midnight.
No. The official script uses uv to install Python 3.11 and Node.js 22. After install, run source ~/.zshrc and hermes doctor to verify.
NodeMini offers dedicated Mac Mini M4 rental by month or quarter. Current pricing is on the rental rates page. LLM API usage (OpenRouter, etc.) is still billed separately by your provider.
A PowerShell install script exists (early beta), but the Dashboard terminal panel works best with WSL2 and the Linux install path. For access and SSH questions, see the Help Center.