The Hermes Agent Field Guide: The Open-Source AI That Writes Its Own Skills

Hermes Agent is Nous Research's open-source agent, MIT-licensed, that runs on hardware you control and keeps state between sessions. The video covered the what. This guide is the how — the commands, the files it writes to disk, and the spots where it doesn't do what the hype says. Everything below is checked against the live docs (linked at the bottom). One thing worth knowing before you start: Hermes is a CLI-first tool that stores everything under a single ~/.hermes/ directory. Memory, skills, scheduled jobs, conversation history — all of it lives there as files and one SQLite database you can read, back up, or move to another machine.

The only thing you install by hand is Git. The installer script provisions the rest for you — uv, Python 3.11, Node.js 22 (for browser automation and the WhatsApp bridge), ripgrep, and ffmpeg (for voice). After it finishes you reload your shell once, then run the setup wizard. --portal is the zero-key path: it logs you into Nous Portal over OAuth, which covers 300+ models on a subscription so you don't paste any API keys to get going.

This is the headline feature, so here's what's literally happening. A skill is a folder under ~/.hermes/skills/ (grouped into category subfolders) containing a SKILL.md file: YAML front-matter on top, a Markdown body below. The body uses plain sections like When to Use, Procedure, Pitfalls, and Verification. A skill can ship supporting folders — references/, templates/, scripts/ — that the agent only opens when it needs them. The trigger is human-in-the-loop, not magic: per the docs, after a complex multi-step task, Hermes will often offer to save the approach as a skill. You say yes, it writes the file using its skill_manage tool. Next session, it sees the skill and reuses it instead of figuring the task out again.

You're not stuck with whatever the agent saves. Skills are just files, so you can list, search, browse the official optional ones, install community skills from the hub at agentskills.io, and edit any SKILL.md by hand. When a skill is wrong or out of date, the fix is to tell Hermes what went wrong and ask it to update that skill — it patches the file with what it learned.

Hermes splits memory into a small always-on layer and a searchable archive. The always-on layer is two short Markdown files in ~/.hermes/memories/: MEMORY.md (the agent's own notes — environment facts, conventions, lessons) and USER.md (your profile — preferences, communication style). The memory docs cap them deliberately small — MEMORY.md around 2,200 characters (~800 tokens) and USER.md around 1,375 characters (~500 tokens) — and inject them into the system prompt as a frozen snapshot at session start (the docs' phrasing: the injection is captured once at session start and never changes mid-session). The archive is your full conversation history in ~/.hermes/state.db, a SQLite database the agent searches with the session_search tool (FTS5 full-text search) when it needs something you discussed weeks ago. It saves proactively, but it's opinionated about what's worth keeping — it skips trivia, raw data dumps, and one-off session noise.

Two things turn Hermes from a terminal toy into something that works while you sleep: the messaging gateway and cron. The gateway connects it to a chat platform — Telegram, Discord, Slack, WhatsApp, Signal, Email, Microsoft Teams, Home Assistant. Message it from your phone, it wakes, runs the task, and replies (voice or text). Cron runs a stored prompt on a schedule and delivers the result to a channel. The schedule field is flexible: a relative delay (30m), an interval (every 2h), a standard 5-field cron expression, or an ISO timestamp.

The marketing around self-improving agents runs ahead of the docs, so here's the grounded version. The skill-creation loop is genuinely useful but it is assisted, not autonomous in the way clips imply: Hermes offers to save a skill after a hard task and you accept it, and stale skills are your job to fix — the docs are explicit that unmaintained skills 'become liabilities,' not that the agent silently retires them for you. It also gives terminal and code-execution tools real power on your machine, so treat tool permissions like you would SSH access. And because it runs on your own infra, uptime, backups, and the model bill are on you — there's no managed SLA behind it.

Once Hermes is answering your phone and running your morning brief, the obvious next thought is: a small business owner would pay for this and never set it up themselves. That's a real gap, and it's a different job from building. Selling agent access means giving each client their own login, metering what they use, putting your brand on it, and billing them — none of which Hermes itself does. That's the layer Knotie handles: a white-label platform where you stand up AI voice and chat agents (Retell, LiveKit, VAPI, Ultravox) plus automations for clients under your own domain, with a customer portal and credit-based billing. It also includes an OpenAI-compatible AI gateway — point the standard OpenAI SDK at a new base_url, hand each client a metered virtual key locked to specific models, add your markup, and the usage bills back to you. Hermes teaches you the craft of running an agent; this is how you'd charge for it.