Most AI tools are web chatbots. You close the tab, they forget you. Hermes Agent is different.

It's an open-source agent runtime from Nous Research — the same lab behind the Hermes, Nomos, and Psyche models. You install it on your own machine, point it at a language model, and it runs in the background continuously. Not a browser tab. Not an IDE plugin. A process that keeps working when you walk away.

What makes it unusual is the range of things it can do once it's running. It connects to over 15 messaging platforms at once — Telegram, Discord, Slack, WhatsApp, Signal, Matrix, email, and more — through a single gateway process. You can message it from your phone in the morning, continue the conversation from your laptop terminal at work, and check results over Slack later. Same agent, same memory, every platform.

It also remembers you between conversations. There's a persistent memory system that records your environment and preferences in small curated files, plus a full session history stored in a local SQLite database with full-text search. On top of that, a user modeling system called Honcho builds a profile of who you are through ongoing interaction. The result is that after a few weeks, the agent knows your projects, your coding style, your preferences — without you having to repeat yourself.

Then there's the skills system. A skill is a reusable set of instructions that teaches the agent how to do something well — deploy to Kubernetes your team's way, write commit messages in your style, review PRs following your conventions. You can install skills from a hub or the agent can create them on its own after noticing a recurring pattern. Skills improve with use and are portable across machines via the agentskills.io open standard.

For execution, Hermes gives you six backends to choose from: local, Docker, SSH, Daytona, Singularity, and Modal. Docker gives you isolation. SSH lets the agent manage a remote server. Daytona and Modal offer serverless persistence where the environment hibernates when idle and costs nearly nothing. You pick the backend per session or per task.

Out of the box you get 47 built-in tools covering file operations, web search, browser automation, shell execution, vision, image generation, and text-to-speech. MCP support lets you plug in external tool servers for GitHub, databases, Notion, and anything else with an MCP server. There's a built-in cron system for scheduled tasks, sub-agent delegation for parallel work, voice mode for spoken interaction, and SOUL.md for defining the agent's personality.

The agent can also spawn isolated sub-agents to handle pieces of a task in parallel, and a feature called Programmatic Tool Calling lets it collapse multi-step pipelines into a single inference call — turning what would be three round trips through the model into one.

It's MIT licensed and free. You only pay for the language model API calls. Provider options include Nous Portal, OpenRouter, OpenAI, Anthropic, and any custom endpoint.

Hermes isn't a polished consumer app. It's a power tool built by people who train models and live in terminals. If you're comfortable with that and want an AI that actually does things instead of just suggesting them, it's worth an afternoon to set up.

The fastest way to start is the learning guide at https://hermesagent101.dev — a 13-step walkthrough from installation to a working agent with memory, skills, and Telegram integration. Takes about 30 minutes.