15% off your workspace - subscribe to our blogNo per-service fees - one plan, unlimited appsDeploy in under 40 seconds99.95% uptime SLAFree tier available - start building today15% off your workspace - subscribe to our blogNo per-service fees - one plan, unlimited appsDeploy in under 40 seconds99.95% uptime SLAFree tier available - start building today
Miget x AIPlansEnterpriseCompareBlogDashboard
Start for Free
Blog/Tutorial/ai/
·

Self-Host Autonomous AI Agents: OpenClaw, Hermes, and Agent Box on Miget

An autonomous agent is not a request-response app. It runs continuously, keeps a workspace, schedules its own tasks, and has a dashboard you check on. That makes it a poor fit for scale-to-zero platforms and a natural fit for a fixed-capacity Resource that is simply always on. deployable.sh has a set of self-hostable agents ready as Compose stacks, and Miget gives each one persistent storage, a private endpoint, and an isolated microVM.

The agents

From the AI agents section of the catalogue:

  • OpenClaw - "an autonomous AI agent you run yourself: web dashboard, scheduled tasks, plugins, persistent workspace."
  • Hermes Agent - Nous Research's self-improving agent that "learns skills from experience, with cron, hooks, and a web dashboard." (Experimental.)
  • Agent Box - "Claude Code, opencode, and Kilo CLI in a persistent web terminal. Start agent runs from anywhere, close the tab, reattach later."
  • Kilo (headless) - "a persistent remote coding agent your IDE extension and CLI clients attach to." (Experimental.)

These are single-service stacks that want about 2 GiB of RAM, so they run on a small paid Resource.

Why self-host them on Miget

  • Always on, by design - a Miget Resource runs continuously. An agent with cron jobs and hooks needs to be up 24/7, not cold-started per request.
  • A persistent workspace - the agent's state, history, and files live on a Miget volume, so "close the tab, reattach later" actually works across restarts and redeploys.
  • Your keys stay yours - these agents call a model provider (Agent Box runs Claude Code, for example). You set your provider API key as an environment variable in your own workspace, not in someone else's SaaS.
  • Private by default - an autonomous agent's dashboard should not be open to the internet. Restrict the app to private access, or put Basic Auth in front of it, from the security settings.
  • Real isolation - each agent runs in its own microVM, so a long-running process with broad permissions is contained.

One honest note on models

Miget runs your apps on CPU compute, not GPUs. That is the right home for the agent itself - the loop, the scheduler, the dashboard, the tools. For the model, these agents call a provider (Anthropic, OpenAI, or your own inference host). You self-host the agent and its state; you bring the model.

Deploy one in three steps

  1. Pick an agent at deployable.sh - say OpenClaw.
  2. Point Miget at it. Miget reads the compose file, deploys the agent, and mounts a persistent volume for its workspace.
  3. Add your keys and lock it down. Set your model provider API key, turn on Basic Auth or private access so only you reach the dashboard, and deploy.

Every push to the tracked branch redeploys the agent, so updating to a new version is one deploy, not a rebuild.

Get started


Self-Host Autonomous AI Agents: OpenClaw, Hermes, and Agent Box on Miget - Miget Blog