Hermes Agent is the new self-hosted AI from Nous Research: an autonomous agent that lives on your server, remembers what it learns, and gets more capable the longer it runs. This is the complete setup, configuration, and operations guide.
Built by Nous Research · 6 messaging channels
Hermes is an autonomous agent process running on your server. You send messages from any app you already use. Hermes picks the right model, executes the work (web search, browser automation, code, file ops, scheduled jobs) and replies. Then it remembers what it learned, generates a reusable skill if it solved something complex, and gets a little better next time.
You (Telegram): "Pull the top 5 LLM reasoning papers from this week and summarize the takeaways"
→ Hermes routes: GLM-4.6 via Nous Portal (long context, low cost)
→ Skill triggers: web_search + page_browse + summarize
→ Auto-curated: stores user preference for paper discovery format
Hermes: "Here are the 5 papers, ranked by citation velocity..."
Total time: 8s · Cost: $0.012 · Skill saved for next time
Commercial AI assistants charge per-seat subscriptions, log every conversation, and lock you to one model family. Self-hosting Hermes flips the equation: you pay for compute, your conversations stay on your hardware, and you can swap models with one command.
| Hermes (self-hosted) | Commercial assistants | |
|---|---|---|
| Cost | $5-30/mo VPS + usage-based model | $20-100/mo per seat |
| Privacy | 100% on your infrastructure | Conversations logged by provider |
| Models | Nous Portal, OpenRouter (200+), OpenAI, Anthropic, GLM, Kimi, MiniMax, custom | Provider's choice, single family |
| Channels | Telegram, Discord, Slack, WhatsApp, Signal, Email, CLI | Depends on plan |
| Memory | Persistent, auto-curated, cross-session | Limited or session-scoped |
| Skills | Auto-generated from your work, agentskills.io compatible | None |
| Sandboxing | Local, Docker, SSH, Daytona, Singularity, Modal | Provider-managed |
| Always-on | Yes, VPS or serverless hibernation | Yes |
Eight features that distinguish Hermes from a chat wrapper or a coding copilot. All open source, MIT licensed, fully under your control.
A single gateway exposes Hermes through Telegram, Discord, Slack, WhatsApp, Signal, Email, and CLI. Voice memo transcription, cross-platform conversation continuity, per-platform access control.
Auto-curated memory between turns. New skills generated after solving complex tasks, then improved as they're used. Full-text search over past sessions with LLM summarization for cross-session recall.
Switch between Nous Portal, OpenRouter (200+), OpenAI, Anthropic, GLM, Kimi, MiniMax, or your own endpoint with one command. No code changes, no lock-in.
Six terminal backends (local, Docker, SSH, Daytona, Singularity, Modal) with container hardening and namespace isolation. Daytona and Modal hibernate when idle, costing nearly nothing between sessions.
Built-in cron scheduler with delivery to any channel. "Send me a Telegram summary of GitHub issues every Monday at 9am": natural language in, scheduled task out, running unattended.
Spawn isolated subagents for parallel workstreams, each with its own conversation, terminal, and Python RPC scripts. Collapse multi-step pipelines into zero-context-cost turns.
Connect any MCP server for extended capabilities. Use the same Model Context Protocol stack you've been building with Claude Code, on a different runtime, on your own infrastructure.
Conversations stay on your hardware. Skills, memory, and config live in ~/.hermes. API keys stay in your .env. No telemetry by default. Full control, full audit trail.
If you already have OpenClaw running, Hermes ships with a built-in migration command. Persona, memories, custom skills, command allowlist, messaging configs, and API keys imported in one shot, with a dry-run preview before anything changes.
Imports SOUL.md, MEMORY.md and USER.md, your custom skills (into ~/.hermes/skills/openclaw-imports/), command approval patterns, platform configs, and allowlisted API keys.
Compare both modules →From first install to a hardened, self-improving agent running across every channel. Lesson 1 is free.
Architecture overview, the learning loop, how it compares to OpenClaw and Claude Code.
Linux, macOS, or WSL2 in under five minutes. Pick a model, run hermes, ship your first task.
Telegram, Discord, Slack, WhatsApp, Signal, Email: full setup with access control and per-platform personas.
Nous Portal, OpenRouter, GLM, Kimi, OpenAI, Anthropic. Routing strategies, fallback chains, monthly budget tracking.
How auto-curation works, when to write skills manually, working with the Skills Hub, and what not to put in memory.
Pick the right terminal backend (local vs Docker vs Modal vs Daytona). Spawn parallel subagents for long-running pipelines.
Scheduled tasks, MCP server integration, monitoring, backups, incident response.
Full hermes claw migrate walkthrough. DM pairing, command approval policies, container isolation, monthly checklist.
Prefer a browser to a terminal? The community-built Hermes WebUI gives you near-1:1 parity with the CLI in a clean three-panel interface. No build step, no framework, vanilla JS.
Community Project
Three-panel layout: sessions, chat, workspace files. Reachable through an SSH tunnel into your VPS, so you keep the same security posture with a friendlier surface. We cover the install in the Pro module. github.com/nesquena/hermes-webui ↗
Start with the free lesson, no signup required. Upgrade to Pro when you're ready for the full setup, model routing, and production ops walkthroughs.
Part of the Fluent course