Hermes & OpenClaw: your in-house agents

Until now, you drive your agent from a terminal, sitting in front of the machine (or over SSH). But once you have an always-on mini-PC, another idea becomes possible: what if you could send it a message from your phone, out in the street, and it got to work? Claude Code does it natively, and that’s where we start. A family of self-hosted agent tools takes the idea further; the two most interesting to know: OpenClaw and Hermes.

Native first: Claude Code can already be driven remotely

Before installing anything, check whether Claude Code is enough. It offers two ways to reach it from your phone, with no third-party software.

Remote Control opens your local session at claude.ai/code or in the Claude app (iOS and Android). You see the conversation, you type, you approve permission prompts, and everything runs on the mini-PC, with your files and your MCP servers.

# an interactive session that can also be driven from your phone
claude --remote-control "mini-pc"

# or server mode: it waits for connections and shows a QR code (space bar)
claude remote-control

In a session that’s already open, /remote-control does the same. You need a Pro, Max, Team or Enterprise subscription, signed in with /login: an API key isn’t enough.

Channels, still in preview, connect a session to a messaging app: Telegram, Discord or iMessage. You write to your bot like to a colleague, Claude Code works on the machine and replies in the same thread. Each channel is a plugin that needs Bun on the machine. For Telegram, once you’ve created a bot with BotFather, the steps happen inside Claude Code: /plugin install telegram@claude-plugins-official, then /telegram:configure followed by the bot token. Then you restart with the channel enabled:

# start Claude Code with the Telegram channel
claude --channels plugin:telegram@claude-plugins-official

Send the bot a message, it answers with a pairing code to confirm with /telegram:access pair <code>, then lock it down with /telegram:access policy allowlist so only your account gets through. Messages only arrive while the session is running: launch it inside tmux (see Working remotely). The full procedure is in the Channels docs, with the Remote Control docs next to it.

The common idea: a bridge between a chat and an agent

OpenClaw and Hermes start from the same principle. You install a gateway (a small service) on the mini-PC. This gateway links a chat channel (typically Telegram) to an agent running on the machine, backed by your local models via Ollama and/or Claude in the cloud.

The result: you write to a bot the way you’d write to a colleague. “Summarize the latest news for me”, “kick off the deployment”, “find where authentication is handled in my repo”. The message arrives on the mini-PC, the agent acts (it has tools, remember the agent loop), and replies. Your workshop in your pocket, with nothing unplugged.

OpenClaw

OpenClaw (first known as Clawdbot) is an open-source agent gateway, MIT-licensed, that you host yourself. Its creator, Peter Steinberger, joined OpenAI in February 2026; since then the project has been stewarded by an independent nonprofit, the OpenClaw Foundation, with no paid tier or hosted service. OpenClaw connects more than twenty messaging apps to a local agent, rich in native commands, and extensible via plugins (with a community catalog, ClawHub, for sharing them): you can teach it to administer your machine, orchestrate several models, trigger your own scripts. It’s the ideal tool to build yourself a tailor-made personal assistant.

Hermes

Hermes Agent (from the Nous Research lab, MIT license) pursues the same idea with its own approach: a command-line agent, paired with a gateway to some twenty messaging apps (Telegram, Discord, Slack, WhatsApp, Signal…), scheduled tasks that deliver their results in the chat, and skills it creates and refines as it goes. It works with any model provider, including your local models via Ollama. It can also hand coding work to another agent: it ships with skills to drive Claude Code, Codex or OpenCode. It overlaps so much with OpenClaw that it offers a migration: hermes claw migrate imports settings, memories and skills from an OpenClaw install.

Concretely, what do you ask it?

It clicks once you see the real uses. A few examples that run great on an always-on mini-PC:

• Monitoring and summaries. “Summarize the last 5 articles from such-and-such RSS feed”: the local model chews through the volume, for free, while you grab a coffee.
• Remote control of your work. “Run the tests on project X”, “deploy the main branch”, “show me the logs from the last 20 minutes”: from your phone, in a meeting or on the train.
• Search in your repos. “Where is authentication handled in my project?”: the agent reads the code and answers, without you opening an editor.
• Reminders and automation. Paired with a scheduled agent, it pings you with a report every morning.
• Idea capture. A note dictated to your bot, and the agent files it, reformats it, adds it to your memory file.

The common thread: anything short, frequent and triggered from the outside is exactly what justifies an always-on machine.

Why the mini-PC is what makes this great

These tools shine precisely because the machine is dedicated and always on:

• Always reachable. Your assistant never sleeps: the mini-PC stays awake while your laptop is closed.
• Private and frugal. Backed by a local model, it can run dozens of times a day with no marginal cost and no data leaving, exactly the hybrid logic: local for volume, Claude for the tough spots.
• Accessible from anywhere. Via your Tailscale private network and a channel like Telegram, you drive it from the other side of the world, securely.

Getting started, carefully