Meet Hermes Agent: An AI That Grows With You

There are a lot of AI tools right now. Most of them are web chatbots that forget you the moment you close the tab. Hermes Agent isn't that.

It's an autonomous agent runtime from Nous Research. The pitch in one sentence: an AI that lives where you do, remembers what it learns, and gets more capable the longer it runs.

That phrase sounds like marketing until you look at the design decisions. Then it starts to look like a different shape of tool than what most people are building.

Not a chatbot

The important thing to understand about Hermes is that it's meant to run in the background, continuously, on a machine you control. Not a tab in your browser. Not a dialog on somebody's website. A process.

Once that process is running, it can do things most AI tools can't:

Take messages from you across a dozen different platforms and reply from the same continuous brain
Run real shell commands inside real sandboxes, including remote ones
Remember everything you told it last week and bring it back into this week's conversation
Schedule its own future work and wake itself up to do it
Delegate pieces of a task to isolated sub-agents in parallel

None of these are hypothetical. They're all in the documented feature set.

Lives where you do

Hermes speaks a long list of messaging platforms out of the box — CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Mattermost, Email, SMS, DingTalk, Feishu, WeCom, BlueBubbles, Home Assistant, and the list is still growing. The messaging gateway is designed so that one agent process shows up on all of them at once.

What this gives you in practice: you can send a Telegram message from your phone in the morning, get a voice note back, switch to your laptop, continue the same conversation in the terminal, then later check the result over Slack. There's no "which app is my AI in right now" — it's the same agent everywhere, with the same memory.

This is unusual. Most AI tools are tied to a single surface. Hermes treats the surface as a detail.

Grows the longer it runs

The second big idea is that Hermes improves itself over time in ways that aren't just "we shipped a new model."

The foundation is a closed learning loop with four pieces that work together:

Agent-curated memory. Hermes keeps persistent notes about your environment and preferences in small Markdown files. The agent periodically nudges itself to write down things worth remembering, rather than waiting for you to tell it.
Cross-session full-text search. Every session is stored in a local SQLite database with an FTS5 index over the content. Weeks later, the agent can find something you said and pull it back into the current context, with an LLM summarization step to keep things tight.
Honcho dialectic user modeling. A conversational user-profile system that builds a picture of you — your habits, your preferences, how you work — through ongoing dialog, instead of treating you like a stranger every session.
Skills that evolve. A skill is a reusable package of instructions — "how we deploy to Kubernetes", "how I like my commit messages", "how this repo does code review." Hermes can install pre-built skills from the Skills Hub, create new ones after noticing a pattern, and refine them over time as they're used.

Skills are compatible with the agentskills.io open standard, which means they're portable across machines and shareable with your team.

Put it all together and you get something that actually gets to know you. The first week it's generic. The third week it knows which repos you care about, what your commit style looks like, and that you prefer pnpm over npm. The tenth week it's solving problems it's already seen before, the way you solved them last time.

Real sandboxing, not just shell access

This is the part that tends to surprise people. Hermes ships with six execution backends:

Local — runs commands on the host machine
Docker — isolated containers on demand, with hardening and namespace isolation
SSH — shell on a remote machine you've given it access to
Daytona — cloud development environments with serverless persistence: the environment hibernates when idle and wakes back up with your working state intact, costing almost nothing while dormant
Singularity — for HPC environments
Modal — serverless containers in the cloud, also with hibernation-style persistence

You pick which one per session, or per task, or per toolset. If you don't want the agent touching your primary machine, give it Docker. If you want it to manage a remote server, give it SSH. If you're running on an HPC cluster, use Singularity. If you want the "costs nothing when idle, wakes up in a second" experience, give it Daytona or Modal. The agent doesn't care — it just picks up a shell and gets to work.

This flexibility is what makes Hermes runnable everywhere from a $5 VPS to a GPU cluster. It's also what makes it safe enough to leave running unattended.

An agent that actually delegates

Most AI tools take a question and return an answer. Hermes can spawn its own sub-agents to work on pieces of a task in parallel — each with its own terminal, its own conversation, its own context — and merge the results back.

It goes further. There's a feature called Programmatic Tool Calling: via the execute_code tool, Hermes can collapse a multi-step pipeline into a single inference call. What used to be "call tool A, get result, call tool B, get result, call tool C, get result" — three round trips through the model — becomes one, with near-zero context overhead. For tasks that need a lot of tool calls (refactoring across dozens of files, say), this is an order-of-magnitude difference.

What's actually in the box

A quick inventory of what you get when you install Hermes:

47 built-in tools — file I/O, web search, browser automation, shell execution, code execution, vision, image generation, text-to-speech, and a lot more
MCP support — plug in any Model Context Protocol server and Hermes picks up its tools automatically
A cron system — describe a scheduled task in plain English and Hermes runs it on its own
Sub-agent delegation — spawn isolated helpers with their own terminals and conversations, useful for parallel work
Voice mode — real-time speech input and output, works in the CLI, Telegram, Discord, and Discord voice channels
SOUL.md — a personality file you write yourself, defining how Hermes talks and behaves
Provider flexibility — Nous Portal, OpenRouter, OpenAI, or any custom endpoint you point at a compatible API

And for the research-minded: trajectory export, batch processing, and reinforcement learning integration with Nous's Atropos.

Built by model trainers

Hermes isn't a side project from an "AI products" company. It's a tool from Nous Research, the open-source model training lab that also shipped the Hermes, Nomos, and Psyche models. They built an agent runtime in part because they needed one for their own work training models.

You can feel this in the product. It's not a polished consumer app with a slick onboarding flow. It's a power tool, built by people who live in terminals.

You'll probably like it if you're comfortable in a terminal, want AI that runs on your own infrastructure, and have ever wished an assistant could actually do things instead of just suggest them. You might bounce off it if you're looking for a one-click experience — configuring a provider, a gateway, and a set of skills takes an afternoon.

For developers, researchers, and anyone building autonomous workflows, that afternoon is an investment that pays back quickly. For everyone else, there's always a web chatbot.

Where to go from here

The fastest way to see if Hermes fits your workflow is to install it and spend twenty minutes playing. The install is a one-line script — no runtime setup, no Python version dance.

Quickstart in your terminal — install, pick a provider, say hi. About fifteen minutes end to end
The full learning path — a structured walkthrough from zero to a working agent with memory, skills, and a messaging gateway
Official docs — the canonical reference, organized by topic

Most AI tools want to be the thing you open. Hermes wants to be the thing that's already there, waiting for you, wherever you happen to be. Whether that's the right shape for you depends on whether you want a tool or a sidekick. If you want a sidekick, it's worth an afternoon.