hermes doctor to verify the stack, hermes setup for your LLM, then hermes gateway install for Telegram 24/7. That is the whole playbook. Below: why Nous Research's MIT agent beats a one-off CLI demo, environment tables, a six-step runbook, BotFather wiring, error fixes, and when a rented Mac Mini M4 is worth it for skill compounding.
Why install Hermes Agent in 2026—and how it differs from OpenClaw
Searching for open-source AI agent tutorials, you will cross Hermes Agent, OpenClaw, LangGraph, and CrewAI. The split is simple: Hermes is a personal on-call operator on your machine—USER.md, MEMORY.md, and skills live under ~/.hermes, with Gateway handling channels. OpenClaw leans harder on multi-channel hubs and ClawHub skills. LangGraph and CrewAI are orchestration frameworks—you assemble memory and deployment yourself.
Self-improving skills: Repeat tasks load from ~/.hermes/skills/. Community reports show skill counts climbing from single digits to a dozen in 30 days, with token use on similar jobs dropping roughly 38%—efficiency signal, not a quality score.
Model-agnostic: Anthropic, OpenAI, OpenRouter, GitHub Copilot, Ollama—all supported. Keys go in ~/.hermes/.env.
MCP extensions: Model Context Protocol servers for GitHub, browser automation, and more.
Install pain point: Passing the CLI once is not the goal—a 24/7 Gateway is. Compare hosts in our three-way scorecard; read how memory compounds in the three-layer memory guide.
Hermes Agent requirements: macOS, Linux, or VPS?
| Item | Minimum | Recommended (production Gateway) |
|---|---|---|
| OS | macOS 12+, Ubuntu 20.04+, WSL2 | macOS 14+ (Apple Silicon) or Ubuntu 24.04 LTS |
| Python / Node | Script installs 3.11 / Node 18+ | Script defaults to Node 22, no sudo |
| RAM | 4 GB (API + Gateway only) | 16 GB (browser automation + local 7B fallback) |
| Disk | ~1.5 GB (incl. skill cache) | ≥ 20 GB SSD |
| Network | GitHub raw + LLM API reachable | Stable, low-latency egress |
| Host scenario | 24/7 uptime feel | Best for |
|---|---|---|
| Laptop lid closed | ≈ 60%, skill compounding breaks | POC only |
| x86 VPS 4 GB | ≈ 99.5%, no macOS-only skill paths | Linux-only users |
| Rented Mac Mini M4 | Datacenter SLA, native launchd | Telegram production, Camoufox automation |
Hermes Agent install: six-step runbook (macOS / Linux)
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash source ~/.zshrc
Headless servers can skip browser deps: bash -s -- --skip-browser. The script installs uv, Python 3.11, Node 22, ripgrep, and ffmpeg, clones to ~/.hermes/hermes-agent/, and adds hermes to PATH. Success line: ✅ Hermes Agent installed at ~/.hermes.
Run the install script: curl command above. Windows has install.ps1, but production still favors WSL2 + Linux script.
Reload your shell: source ~/.zshrc or source ~/.bashrc fixes hermes: command not found.
Health check: hermes doctor validates Python, Node, ripgrep, ffmpeg, PATH, API keys—about 12 checks. hermes --version should show v0.14.x or v0.15 candidate.
Configure the brain: hermes setup wizard, or hermes model for OpenRouter / Anthropic / Ollama alone. Keys land in ~/.hermes/.env, e.g. hermes config set OPENROUTER_API_KEY sk-or-xxxx.
CLI smoke test: run hermes, try "summarize today's AI news" or "count files in this directory"—confirm tool calls and model routing.
Upgrade and migrate: day-to-day hermes update; if config errors appear after upgrade, run hermes config check && hermes config migrate.
Hermes Agent Telegram setup: BotFather to gateway install
CLI alone means local chat only. For phone-side tasking, you need Gateway. Flow: find @BotFather on Telegram → /newbot → save the token → use @userinfobot for your numeric user ID.
hermes gateway setup hermes gateway install hermes gateway start hermes gateway status
The wizard writes TELEGRAM_BOT_TOKEN and TELEGRAM_ALLOWED_USERS. On Linux production, sudo hermes gateway install --system plus journalctl -u hermes-gateway -f for logs; macOS uses a launchd user service.
| Error / symptom | Cause | Fix |
|---|---|---|
hermes: command not found | PATH not loaded | Reopen terminal or source rc file |
API key not set | No LLM configured | hermes model |
| Group chat silent | Bot privacy mode | Disable Group Privacy in BotFather, kick and re-add bot |
ModuleNotFoundError: dotenv | System Python used | Use ~/.hermes/hermes-agent/venv/bin/hermes |
Tip: Also useful: hermes tools, hermes config check. Declare mcp_servers in ~/.hermes/config.yaml; on headless hosts, hermes mcp login plus SSH tunnel handles OAuth.
After install: skill compounding and Mac Mini M4 hard data
Star scale: As of June 2026, hermes-agent sits near 174,720 GitHub stars. Install and Telegram docs iterate fastest on the macOS path.
Power draw: M4 idles around 4–6 W; Gateway plus automation typically 15–25 W—fine for always-on duty.
Unified memory: On 16 GB UMA, Gateway residency is roughly 4 GB class, sharing address space with Camoufox and SQLite WAL—less swap thrash than a small x86 VPS.
Closed laptop lids, Raspberry Pi I/O limits, and Linux VPS without macOS browser skills are not install failures—they are wrong host choices that break Gateway SLA. If you want Hermes on Telegram 24/7 with skills that keep writing themselves, MESHLAUNCH Mac Mini M4 cloud bare-metal rental is usually the cleaner path: same curl install over SSH, launchd supervision, and you can tarball ~/.hermes before offboarding. Check rental pricing for a day trial; ops boundaries live in the help center.
No. install.sh uses uv for Python 3.11 and Node 22. If doctor fails, confirm you are calling venv hermes, not system Python.
After disabling Group Privacy in BotFather, remove the bot from the group and re-invite it. Verify TELEGRAM_ALLOWED_USERS. You can order a dedicated M4 to isolate Gateway testing.
Disk memory persists, but an offline Gateway stops skill compounding and Cron. Production needs hermes gateway install on a 24/7 host—see the memory architecture guide.