1. What Is Hermes Agent?
Hermes Agent is an open-source, self-improving AI agent framework built by Nous Research, released under the MIT license. Tagged "the agent that grows with you," it is an autonomous assistant that plans tasks in natural language, calls tools, executes code in sandboxes, and — its signature trick — writes its own skills as it solves problems, so the same task is faster and cheaper the second time. If the term is fuzzy, it's worth pinning down what makes an LLM "agentic" in the first place — Hermes is a textbook implementation of those properties.
Unlike a chatbot that lives in one tab, Hermes runs as one agent with one memory across every surface you use: a terminal CLI, a native desktop app, and messaging gateways for Telegram, Discord, Slack, WhatsApp, Signal, and email. Sessions, configuration, API keys, skills, and memory sync across all of them.
Hermes Agent at a glance:
- • License: MIT — audit it, self-host it, fork it
- • GitHub: ~219k stars, ~41k forks, 81% Python / 16% TypeScript
- • Current version: v0.19.0 (desktop app shipped June 2, 2026 as public preview)
- • Model-agnostic: OpenAI, Anthropic, OpenRouter, or 300+ models via Nous Portal
- • Persistent memory: projects, solution paths, and full-text session search (FTS5)
2. Hermes Agent vs. Hermes LLM vs. OpenHermes — Which Is Which?
The name "Hermes" is overloaded, and search results mix three different things (four, if you count the French fashion house Hermès — that one does not ship a CLI):
- • Hermes Agent — the agent framework covered in this article: planner, tools, skills, memory, gateways. It can be powered by any capable LLM.
- • Hermes (the LLM family) — Nous Research's line of fine-tuned open-weight language models (Hermes 2, Hermes 3, and successors), known for strong instruction following and tool calling.
- • OpenHermes — the earlier open dataset and fine-tune lineage that put the Hermes name on the map in the open-source community.
The agent is deliberately not tied to Nous's own models. You point it at any compatible endpoint with hermes model — a design choice that separates orchestration quality from model quality, and one reason engineers take the codebase seriously.
3. Architecture: What's Actually in the Repo
The hermes-agent GitHub repository is organized the way you would design an agent platform from first principles:
- • agent/ — core planning and execution loop
- • skills/ — procedural memory: skills the agent writes and reuses
- • tools/ — 40+ built-in tools (browser, search, images, TTS, files)
- • providers/ — LLM provider integrations
- • gateway/ — Telegram, Discord, Slack, WhatsApp, Signal bridges
- • hermes_cli/ — full TUI with slash-command autocomplete and history
- • plugins/ — extension system, compatible with the agentskills.io standard
Two architectural decisions stand out. First, subagent delegation is process-level: a delegated subagent gets its own conversation, its own terminal, and Python RPC — closer to spawning a junior engineer than calling a function. Second, execution is sandboxed by default with pluggable backends — a design that matters more than ever after OpenAI's models escaped a sandbox and hacked Hugging Face:
| Sandbox Backend | When You'd Use It |
|---|---|
| Local | Fastest iteration on your own machine, no isolation |
| Docker | Container isolation for untrusted or repeatable tasks |
| SSH | Run the agent's hands on a remote box you control |
| Singularity | HPC and research cluster environments |
| Modal / Daytona | Serverless cloud sandboxes that scale to zero |
4. The Self-Improving Loop: Skills as Procedural Memory
Most agent frameworks treat every task as a cold start. Hermes's built-in learning loop is different: when the agent solves a problem, it can distill the solution path into a named, reusable skill. Next time a similar task appears, the skill short-circuits the exploration phase — fewer tokens, fewer wrong turns, faster completion.
Combine that with persistent project memory and cron-style scheduling ("send me a morning briefing, back up this folder nightly") and you get something closer to an accumulating co-worker than a stateless assistant. This is the pattern we expect most serious agent stacks to converge on — the interesting part is watching it work in a codebase you can read.
5. How to Download Hermes Agent (Mac, Windows, Linux)
The official download source is hermes-agent.nousresearch.com — no account gate, no waitlist. The Hermes Desktop app (public preview since June 2026) gives you streaming tool output, a side-by-side preview pane, a file browser, voice input/output, and a settings UI — no terminal required. CLI and desktop share the same agent core and memory.
| Platform | Method | Install Command |
|---|---|---|
| macOS 12+ | DMG installer or terminal | curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash |
| Windows 10/11 | EXE installer or PowerShell | iex (irm https://hermes-agent.nousresearch.com/install.ps1) |
| Linux / WSL2 | Terminal | curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash |
Standard caution applies to piping curl into bash: read the install script first, or install from the GitHub repo if your threat model demands it. After install, run hermes to start the TUI and hermes model to pick your provider.
6. Models, Nous Portal, and Pricing
The framework itself is free and open source. Bring your own API keys (OpenAI, Anthropic, OpenRouter, or any compatible endpoint), or use Nous Portal — Nous Research's hosted gateway to 300+ models with tiered subscriptions (Free, Plus, Super, Ultra) that bundle monthly credits. Self-hosters pay nothing beyond their own inference.
7. An Engineer's Take: Why This One Matters
Plenty of agent frameworks demo well. Three things make Hermes worth your reading time as an engineer:
- • The memory model is the product. Skills-as-procedural-memory plus searchable session history is a concrete answer to the "agents forget everything" problem — in shipping code, not a paper.
- • Surfaces are thin, the core is one. Desktop, CLI, and six messaging gateways over a single agent core is the right decomposition; most stacks get this wrong and fork per-platform.
- • Sandbox pluggability. Local-to-Modal execution behind one interface means the same agent scales from laptop experiments to serverless production without rearchitecting.
If you are building your own agent stack, read skills/ and gateway/ before writing another orchestration layer from scratch.
8. Frequently Asked Questions
What is Hermes AI?
"Hermes AI" usually refers to Hermes Agent, Nous Research's open-source autonomous AI agent — though it can also mean the Hermes family of fine-tuned LLMs from the same lab. The agent is the framework; the LLMs are models that can power it.
Is Hermes Agent free?
Yes — MIT-licensed and free to download, self-host, and modify. You pay only for the model inference you use, either via your own API keys or a Nous Portal subscription tier.
Does Hermes Agent run on Windows?
Yes. Windows 10/11 gets a native EXE installer for the desktop app, or a one-line PowerShell install for the CLI. WSL2 is also supported via the Linux script.
Where are the official website and documentation?
The official website is hermes-agent.nousresearch.com, documentation lives at /docs, and the source is at github.com/NousResearch/hermes-agent.