A fully self-hosted, private stack
“I want Hermes running entirely on my own hardware with local models and no data leaving my network”
"Self-hosted" quietly breaks at whichever layer you forget. People containerize the agent, run a local model, and then wire in a cloud memory provider — and the private stack has a cloud dependency in the one place that sees everything. So this bundle is deliberately ordered by where the leaks actually happen: runtime, then inference, then memory. The safety layer is not optional just because the network is yours; the credential and PII exposure is from the agent to itself. Everything here has a heavier local-hardware cost than the cloud equivalent, and that tradeoff is the whole point.
Real posts from the Hermes community, mirrored from Nous Research's user stories corpus.
I set up my Hermes Agents with local SearXNG (in a container that they share) before ddgs (duckduckgo) was added, because I thought all API use of DDG was paid. I've been pretty happy with searxng so far.
Users want secure remote access to the Hermes API server / Open WebUI without exposing ports publicly. Tailscale serve provides zero-config HTTPS tunneling over a private mesh.
I run Hermes self-hosted on a single edge-class GPU with a 4B Gemma model. I work with legal-domain material and internal systems I cannot ship to third-party APIs. Self-hosting the main loop is non-negotiable.
A persistent agent that runs on your own infrastructure, reaches you across messaging platforms, and gets more capable with use. Supported platforms include Linux, macOS, WSL2, and Termux — start on your laptop, continue on Telegram.
I'm using Hermes. The same applies to all agents, don't give it free reign over all your stuff. Run it within a sandbox. https://github.com/nousresearch/hermes-agent
Running a 4-agent Hermes system 24/7 on a 32GB Ubuntu laptop, managed by systemd + custom watchdog. default (PM) — Feishu WebSocket; dev (Developer) — ACP delegated; devops (Ops) — ACP delegated; ip-agent (Content) — Cron scheduled. 5 MCP Servers, 34 tools…
- Local models are meaningfully weaker at long tool-use chains. Expect to simplify workflows that worked on a frontier model.
- Sandboxing and remote access pull in opposite directions — decide your access story (tunnel, not open port) before you harden the runtime.
- The catalog has no strong self-hosted, local-first web-search option yet — community stories reach for a self-run SearXNG container, which isn't in Atlas as a Hermes-specific project.