Two years ago, an "AI agent" meant a browser tab you paid a monthly fee to use. Today you can run one that lives on your own server, answers you in Telegram or Slack, uses your Anthropic or OpenAI key, and never routes your data through someone else's product. Two open-source projects lead this shift: OpenClaw and Hermes.
This guide explains what each one is, how they differ, and — the part most articles skip — what it actually takes to run one in production.
What is a self-hosted AI agent?
It's a small program — a "gateway" — that you run on a machine you control. It connects to your messaging apps on one side and to a large language model on the other. You bring the model's API key, so you pay the provider directly at their published rate. Nothing sits between you and the model except software you can read.
That single design choice is why people search for these tools: your prompts, files, and chat history stay on your infrastructure, and your cost is metered by tokens, not by seats.
OpenClaw: the proactive personal agent
OpenClaw is a free, MIT-licensed, self-hosted personal AI assistant. You run one gateway process; it bridges a long list of channels — Telegram, Slack, Discord, WhatsApp, Signal, iMessage, Matrix, Microsoft Teams and more — to an always-available agent, with a local browser dashboard for chat and configuration.
It's model-agnostic by design: point it at Anthropic (Claude), OpenAI, Google, DeepSeek, or a local model through Ollama or LM Studio. The docs deliberately don't prescribe a provider — you supply the key for whichever model you trust.
Its standout feature is a heartbeat: a scheduled, cron-like loop that lets the agent act on its own without being prompted — checking things, following up, running routines. That's the line between a chatbot and an agent. (NVIDIA and DigitalOcean have both published technical write-ups on running OpenClaw locally, if you want an outside view: NVIDIA, DigitalOcean.)
Hermes: the agent that learns
Hermes Agent, from Nous Research, is also open-source (MIT) and self-hosted, and it's frequently framed as the OpenClaw alternative. Its distinguishing idea is a built-in learning loop: it turns things it does into reusable "skills," keeps persistent memory, runs scheduled tasks, and can spin up subagents. Its tagline — "the agent that grows with you" — is the pitch. You talk to it from a terminal or in gateway mode from your chat apps, and manage it from a web dashboard. Like OpenClaw, it's bring-your-own-key across OpenAI, Anthropic, DeepSeek, OpenRouter and custom endpoints, with no lock-in.
OpenClaw vs Hermes, side by side
| OpenClaw | Hermes Agent | |
|---|---|---|
| Made by | Peter Steinberger (open source) | Nous Research |
| License | MIT | MIT |
| Best known for | Proactive "heartbeat" + huge channel list | Self-improving learning loop + memory |
| Interfaces | Chat channels + web dashboard | Terminal + gateway + web dashboard |
| Models | Bring your own key — Anthropic, OpenAI, Google, DeepSeek, local (Ollama/LM Studio). No lock-in. | |
| Data | Stays on your server; only the model call leaves, to the provider you chose. | |
Why self-host instead of a SaaS agent?
The honest answer: control and cost structure. Here's the trade, without the hype.
- Data ownership. Conversations and files stay on your box.
- Real per-token pricing. Pay the LLM provider directly — no platform markup.
- No per-seat fees. Cost scales with usage, not headcount.
- No lock-in. Swap models or providers whenever.
- Custom channels a SaaS product may never support.
- Uptime and updates. The server is yours to keep running.
- Security. An agent that can run commands is a surface to sandbox.
- The LLM bill. Still metered — set provider spend caps.
- Setup. Someone has to deploy and maintain it.
The catch: someone has to run it
Installing either agent is a few commands. Running one properly is more: a server, a domain, an HTTPS certificate that auto-renews, the container kept alive across reboots, config and secrets managed safely, and updates applied without breaking things.
That's the gap HostAiHub closes. You pick an agent, we auto-install it to your own private, SSL-secured subdomain, wired to your LLM key — usually live in a couple of minutes. You still own the instance and the data; you just don't babysit the plumbing.
Which one should you pick?
- Choose OpenClaw if you want a personal assistant across many chat apps that also acts on a schedule.
- Choose Hermes if you want an agent that accumulates skills and memory over time.
- Either way, you keep your data and your choice of model. That's the whole point of self-hosting.
Both are strong, actively developed, and genuinely open source. The right call depends on whether you value OpenClaw's proactive reach or Hermes's learning loop — and you can deploy either one without touching a server config file.