Hermes Agent

What it is

Hermes Agent, from Nous Research, is the odd one out on this board. Everything else here is pointed at a repository and aimed at developers. Hermes is a Python service you install on a machine you control, and it is pointed at a business: a mailbox, a form, a queue, a chat channel. Something arrives, the agent decides what it is, and it acts — replies, files it, raises a ticket, updates a sheet.

The two properties that matter are both in the spec above. It is self-hosted, so it runs whether or not anyone's laptop is open and your data stays on infrastructure you chose. And it is provider-agnostic, so the model is a line in a config file rather than a purchase decision baked into the tool.

How the loop works

A trigger, a decision, an action, a record. Inputs come from connected channels — it speaks to a long list of messaging platforms, which is why nobody at the business has to learn a new interface; the agent becomes another contact in the tool they already have open. The loop underneath is the ordinary one: model call, tool call, result, repeat, with MCP as the route to whatever systems you already expose.

Its distinctive feature is a learning loop that turns things it has worked out into reusable skills, with a community skills marketplace alongside it. Take that at the right altitude. It does not get quietly cleverer about your business overnight. It means a task solved once need not be re-derived every time — which also means your early corrections are the ones that get crystallised, and your early inattention is too. Budget real supervision for the first fortnight.

Isolation and permissions

The isolation field reads "your own host", which is the honest answer and not a reassuring one. There is no vendor sandbox between the agent and the machine. What that machine can reach, the agent can reach: the inbox it reads, the systems it writes to, the credentials in its environment. Run it on a dedicated VPS or container rather than an office desktop, and give it its own scoped accounts — a mailbox user, an API key with the narrowest permissions that still let it work — so that revoking access later is one action rather than an investigation.

Then consider the input. An agent that reads a public mailbox is an agent reading text written by strangers, which can carry instructions. Keep the irreversible actions — sending money, deleting records, emailing clients unsupervised — behind explicit approval, and log every action somewhere a human reads on Monday morning.

Who it is for

Small businesses with a repetitive intake problem and no developer on staff, or a friendly one nearby for a day's setup. Enquiry triage, appointment chasing, invoice and document filing, first-line support — work that is high-volume, low-judgement and currently done badly at 6pm by whoever is left. It is not a coding agent, and using it as one would be a mistake.

Limits

• You are the operator. Uptime, backups, updates and key rotation are yours. A hosted alternative exists, and that is the trade you would be making.
• Three separate costs. Hosting, model usage and your own attention. Only the first is fixed, and the third is the one people forget.
• Provider-agnostic means quality varies. A cheap model on a nuanced client email is a false economy; start cheap, but measure before you settle.
• Self-improvement cuts both ways. Habits learned from unreviewed work are harder to unpick than they were to acquire.
• Small project, fast-moving. Installation steps and integrations change; the docs are the authority and any guide, including ours, ages.

Alternatives on this board

• Managed Agents — if you would rather someone else ran the host and the sandbox.
• Claude Agent SDK — if you have engineers and want to build the intake agent yourself.
• CrewAI or LangGraph — when the work is a defined multi-step process rather than an open-ended inbox.
• OpenClaw — the personal-scale counterpart: self-hosted and model-agnostic, pointed at one person's machine rather than a business.