Agents · Nous ResearchFor technical users

Hermes Agent

A self-hosted autonomous agent that runs on your own infrastructure, remembers you across sessions and writes its own reusable skills. Powerful, and genuinely technical to run.

Start here

What this helps you do: Understand, hands-on, what an always-on agent with memory can do, if you are technical and comfortable running your own server.
What you leave with: A practical way to use Hermes Agent on real work, plus the human checks before you rely on it.
What to check before trusting it: It runs real commands and acts autonomously, so its permission system is essential and a misconfigured agent on a networked server is a genuine security risk.

Ask Bit to set up a workflow in Hermes Agent

At a glance

Version at last check: Hermes Agent v0.22.2, released 15 October 2026; desktop and command-line installation options vary by platform.
Category: Agents
Pricing at last check: The software is free under the MIT licence; hosting, model and tool use can cost extra. Nous Portal lists Free, Plus US$20/month, Super US$100/month and Ultra US$200/month. Check included credits and the selected provider before budgeting.

What it does

Persistent memory and a learning loop

Keeps memory across sessions and turns experience into reusable skills it refines as it goes, building a picture of your recurring work over time.

For work An assistant that can reuse saved information and skills for recurring tasks; check whether that improves your results.

Runs unattended on your own hardware

Installs on Linux, macOS, WSL2, native Windows, Nix or Android, and runs as a persistent process on infrastructure you control, from a small server to a GPU machine or a serverless sandbox that sleeps when idle. Scheduled tasks run on a built-in cron (a timetable that runs jobs without you).

For work Standing automation that continues without you sitting there.

Lives where you chat

Connects through a gateway to more than twenty messaging platforms, including Telegram, Slack, Discord and WhatsApp, and has a real-time voice mode.

For work Driving the agent from whatever app you already use, by text or voice.

Model-agnostic, with a tool gateway

Works with OpenAI, Anthropic, Google, a local model or Nous's own portal, and plugs in external tools and servers through MCP, plus web search, browsing and image generation.

For work Building a private agent that is not tied to one provider.

Permission and approval layer

Execution permissions can block actions or require approval. Configure and test those controls before allowing unattended commands.

For work Keeping a hand on an agent that can act on your systems.

Not for

People who want fully managed setup and operations. Desktop installers simplify access, but you still need to configure a model, permissions and running costs.

Real work with it

A developer or self-hoster

Standing personal automation on your own server

Deploy it on a server you control and connect it through Telegram or Slack.
Let it accumulate skills for recurring chores like triage or research digests.
Use the scheduler for tasks that should run on their own.

Human checkpoint The approval layer exists for a reason: gate anything that spends money, sends an external message or writes to production before it can run unattended.

A privacy-conscious technical user

A private agent stack you own end to end

Point it at your own model endpoint and wire in the tool servers you trust.
Use subagents to run parallel jobs.
Review the skills and memory it builds before relying on them.

Human checkpoint You own the operations, the security and the bill; decide the model, the budget and what the agent is never allowed to do.

The part that stays yours

An agent that runs unattended and keeps information between sessions needs an owner, boundaries and an off switch. What it may touch, what it may do without asking, and who answers for what it does are the human decisions that make it safe.

When to pick something else (3)

A hosted assistant's agent mode (ChatGPT, Claude, Gemini)

You prefer a hosted service to maintaining the agent infrastructure yourself. Compare permissions, privacy, setup and review effort.

Manus

You want to hand off a long task to the cloud and walk away, without running any infrastructure yourself.

Genspark

You want a polished, managed 'super-agent' platform for documents and multi-step work, hosted for you.

Watch-outs (4)

It runs real commands and acts autonomously, so its permission system is essential and a misconfigured agent on a networked server is a genuine security risk.
'Free open-source software' still means paying for model calls and hosting, and an autonomous agent that spawns subagents can run up token cost quietly.
It is not for non-technical users: setup assumes comfort with a terminal, containers and API keys, and the quick-install claim understates configuring a model, tools and safe permissions.
Model access and hosting have separate costs. Check the selected provider or portal offer rather than treating the open-source licence as a promise of free operation.

Sources

Checked: 17 October 2026
Review due: 31 October 2026
Owner: The Human Bit editorial