Hermes Agent FAQ

Hermes Agent is an MIT-licensed, self-hosted AI agent from Nous Research. The core difference from ChatGPT / Claude: Hermes is a persistent agent that lives on your server, with durable memory and a Skills system. A regular chat UI starts from zero every session; Hermes remembers what you said yesterday, packages multi-step tasks into reusable skills, and gets sharper the more you use it.

No, this is not official. This site is a community-maintained Chinese + English guide, not affiliated with Nous Research. Official docs live at hermes-agent.nousresearch.com. Everything here is tested by the author against the official releases; when in doubt, the official docs win.

Yes — MIT license, source on GitHub / NousResearch / hermes-agent. Commercial use, modification, and redistribution are all allowed as long as the original notice is preserved. The tool itself is free — you only pay for the LLM provider you pick (OpenAI / Anthropic / OpenRouter / DeepSeek / etc.). Local models cost nothing.

Similar DNA, different philosophy. OpenClaw wraps an agent around a messaging gateway (breadth-first); Hermes wraps a gateway around a learning agent (depth-first). Two big differences:

• Skills — OpenClaw skills are static, human-authored templates. Hermes distills reusable skills automatically from completed tasks.
• Memory — OpenClaw stores plain Markdown. Hermes uses SQLite + FTS5 full-text search + LLM summarization, a four-layer memory system.

Official one-shot migration: hermes claw migrate imports SOUL.md / MEMORY.md / skills / API keys. See the migration guide.

Native Windows is not supported — you need WSL2. Officially supported: Linux / macOS / WSL2 / Android (Termux). One-liner installer:
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
Don't prefix with sudo — the installer writes to ~/.local/bin; sudo pushes it to /usr/local/bin and breaks path resolution. If you did, run sudo rm /usr/local/bin/hermes first, then reinstall. Android users replace .[all] with .[termux]. Full install walkthrough: Day 1.

Depends on where the LLM runs:

• Cloud APIs (OpenAI / OpenRouter / DeepSeek) — a $5 VPS is plenty. All compute is remote.
• Fully local — 32B+ parameter models, 24 GB+ VRAM minimum.

Hard requirements: Python ≥ 3.11; the chosen model must expose at least 64K tokens of context — anything smaller is rejected at startup (multi-step tool calls won't fit). Model selection deep dive: Day 2.

Everything OpenAI-compatible works: OpenAI, Anthropic, Google, OpenRouter (200+ models), Nous Portal, NVIDIA NIM, Hugging Face, plus Chinese providers like DeepSeek, Kimi (Moonshot), z.ai/GLM, MiniMax, Qwen, Xiaomi MiMo. If the endpoint implements /v1/chat/completions, point hermes model at it. Switching providers doesn't need a restart — memory and skills carry over.

Yes. Start a local inference server (Ollama / vLLM / llama.cpp / SGLang), then run hermes model → select "Custom endpoint" → enter http://localhost:11434/v1 (Ollama's default). Zero traffic to public APIs. Pick a 32B+ model that natively supports 64K context.

The tool is free; cost comes from LLM usage:

• Local models — $0/mo
• Chinese APIs (DeepSeek / Kimi / GLM) — $1–5/mo light use
• Claude / GPT — $3–15/mo light, $15–70/mo daily deep use
• OpenRouter — per-model pricing, swap cheap-to-expensive on demand

Savings tip: default to cheap Chinese models; swap to Claude for hard tasks. hermes model switches instantly; memory stays. Multi-provider routing: Day 2.

Built-in Messaging Gateway supports Telegram, Discord, Slack, WhatsApp, iMessage, WeChat (itchat), and WeCom (enterprise). Basic recipe:

1. Enable the channel in ~/.hermes/config.yaml (e.g., messaging.telegram.enabled: true)
2. Configure an allowlist (only respond to specific users / channels)
3. Run hermes gateway run and follow the prompts

Full walkthrough: Day 3 · Go Social. Note: each profile needs its own bot token — profiles can't share one.

Both are persistent across sessions, but with distinct roles:

• Memory (factual) — facts about you and your projects: your name, preferences, ongoing work, anything mentioned in past chats. Auto-retrieved by relevance.
• Skills (procedural) — step-by-step how-to for recurring tasks: "deploy Next.js to Vercel", "draft weekly report", "debug nginx 502". Auto-invoked on similar tasks.

One-liner: Memory is "what you know"; Skills is "what you can do". Both survive model switches. Memory hands-on: Day 6; Skills hands-on: Day 5.

API calls only hit the LLM provider you configure — Nous Research collects no telemetry. Conversations, memory, and skills live locally in a SQLite DB under ~/.hermes/.

• Want full offline? Use a local model (see Q above).
• Need encrypted backups? Just tar ~/.hermes/.
• Enterprise compliance? Point Hermes at a private OpenAI-compatible endpoint (self-hosted vLLM) — data stays on your network.

Built-in tools and endpoint configuration: Day 4.

The installer writes hermes to ~/.local/bin/, which isn't in every shell's default PATH. Two steps:

1. source ~/.zshrc (or ~/.bashrc), or just open a fresh terminal.
2. Still missing? Run echo $PATH. If ~/.local/bin isn't in it, append export PATH="$HOME/.local/bin:$PATH" to your profile.

Install pitfalls collected: Day 1.

Three tools:

• hermes doctor — the official diagnostic command: checks deps, config, model connectivity, and prints exactly what to fix.
• ~/.hermes/logs/ — rotating logs. Gateway issues are usually visible here.
• Official FAQ & Troubleshooting — covers WSL systemd issues, WhatsApp disconnects, MCP tools not showing, Telegram log spam.

For 429 rate-limit: wait, upgrade your plan, or switch providers via hermes chat --provider <alt>. Full automation and diagnostics: Day 7.