A Practical Guide to Setting Up Hermes as an AI Assistant

Peter Yang presents a practical course for configuring Hermes as a personal AI assistant. He covers installation, messaging, voice, Google Workspace, and scheduled routines. His recommendations, model choices, and price examples reflect the time of the video; check the current product documentation and account terms before configuring your own setup.

The first choice is where the agent should run. Peter compares a dedicated Mac, an everyday laptop, and a rented cloud server. A machine that is asleep or offline cannot reliably run local scheduled jobs, while a server requires maintenance. He prefers a separate Mac user and email account so he can control what the assistant can access.

Peter installs the desktop application, connects a model provider, and checks that the intended model and effort setting are actually selected. Model availability, subscription eligibility, and costs can change, so the choice should be checked against the current provider terms and usage limits.

To connect Telegram, he creates a bot with BotFather, places its token and his permitted Telegram user ID in Hermes, restarts the gateway, and sets a home channel for scheduled messages. The bot token is a credential: keep it private, restrict who can use the bot, and rotate it if exposed.

Peter explains the basic flow: a message enters through the gateway, the model plans a response, and the agent may call tools to retrieve information or make changes. Memory and skills can preserve useful preferences and workflows, but users should inspect what is stored and be deliberate about sensitive information.

For personalization, he asks the assistant questions about his preferences and reviews the user and personality files it creates. For troubleshooting, he demonstrates diagnostic and gateway-restart commands and suggests consulting the documentation or a coding assistant. Review proposed fixes before allowing them to change the system.

Peter enables voice replies through the messaging interface and discusses voice selection. He also demonstrates suppressing some system messages and routinely approving permissions. Those are his preferences; keeping visibility into actions and reviewing each permission can help users understand what their agent is allowed to do.

The Google Workspace walkthrough creates a cloud project, enables relevant APIs, configures an OAuth application and a test user, and completes authorization. Peter says his normal setup uses a separate account, read access to email and calendar, and selective access to files. In the demonstration he grants broader permissions to test features, so the precise access actually granted should be checked rather than assumed.

He tests an inbox summary, creates a calendar event after confirmation, and asks the agent to update a document. The first document edit disrupts formatting, and he asks for a correction. This illustrates why users should verify changes and limit write access to the specific resources and actions they intend to permit.

Other examples include WhatsApp or Slack, meeting notes, GitHub backups, and YouTube research. Peter suggests a private repository for backing up the setup; before doing so, exclude or securely manage bot tokens, OAuth files, and other secrets. Availability of transcripts and integrations varies.

For recurring work, Peter recommends a three-step loop: create a skill, run it manually and refine the result, then schedule it. His examples include weekend activity ideas, a morning briefing, and business or health summaries. Check sources and outputs, especially for sensitive or consequential decisions, and make sure the host is running when the job is due.