Hermes Agent

What it actually does

Hermes Agent is an AI assistant you install rather than log into. It lives on your own machine or a VPS, keeps memory across sessions, writes its own reusable skills as it works, and reaches you through Telegram, Discord, Slack, WhatsApp, Signal, email or a terminal. It can run shell commands, edit files, drive a browser, search the web and schedule jobs that fire while you sleep.

The part that separates it from a chat window is continuity. Ask it to check something every Monday and it keeps doing that, remembers what it found last week, and messages you the result. It also hands work to subagents with their own context, which keeps a long research job from flooding the main conversation.

There are three ways to run it and they matter more than the feature list. The terminal for real work, a desktop app on macOS or Windows when you want a window instead of a shell, and one-click cloud hosting if you would rather not leave a machine awake. We run ours through Telegram, which sounds like a gimmick and is not: an agent you can message from a phone while standing in a queue gets used, and one that lives in a terminal tab gets forgotten.

It is MIT-licensed. Nous Research raised money on it, and the code is still yours to read, fork and run without asking anyone.

What it costs

Nothing, if you self-host and bring your own model API keys. That is the honest entry price and it is the one we would recommend starting from.

Nous Portal is the paid convenience layer. Free is $0 and limited to free models. Plus is $20 a month for $22 in credits with a $10 rollover cap. Super is $100 for $110 in credits, and Ultra is $200 for $220, both with bigger rollover caps. Hosted tools and cloud agent hosting bill against the same balance.

Read that twice, because the word subscription does a lot of work here. The $20 is not a seat. It is $22 of metered credits spent per token at published per-model rates, and an agent that chains tool calls across a long task spends them faster than a person typing questions does. One user on the Hermes subreddit described a single kanban cleanup task burning roughly $6 of a $20 balance, and another reported a dollar gone on a few test prompts while setting things up. Neither was doing anything exotic.

What nobody tells you at signup is that the bill is mostly a model-choice problem rather than a product problem. Point it at a frontier model and a long task costs real money. Point it at a cheap or free model for the routine jobs, and the same $20 lasts the month. Switching is one command, and deciding which model handles which job is the single biggest lever you have over the bill.

If your mental model is a flat monthly assistant plan, this will feel wrong in week two. If your mental model is metered API billing with a nicer front door, it is priced about where it should be, and the rollover caps are the detail to watch since unspent credit above the cap does not carry.

Where it falls down

The evidence base is thin and we are not going to pretend otherwise. There is no G2 listing, no Capterra profile, no Trustpilot score. What exists is a GitHub repository with around 246,000 stars and 51,500 forks, an active subreddit created in March 2026, and press coverage of the funding rather than of the product. TechCrunch reported the GitHub following at roughly 214,000 stars in July 2026, so the growth is real and fast, but stars measure interest, not satisfaction. We scored this on a much smaller pile of evidence than most pages on this site.

Setup is the second problem. Installing is one command, but running it well means API keys, a machine that stays awake, and enough comfort with a terminal to debug a config file. The desktop app softens that. It does not remove it.

Then there is the security question, and it is not theoretical. This is software you deliberately give shell access, file access and your credentials. That is the whole point, and it also means a bad instruction from a web page it reads is a real risk to think about before you hand it your home directory.

Per-token pricing through Portal also runs above going direct in some cases, which people notice and say out loud in the community.

Who should buy it

Developers and technical people who want an agent they own. If you already pay for model API access, self-hosting costs you nothing extra and you get memory, scheduling and chat-app delivery for free. That is a genuinely good deal and it is why this is rated where it is.

Anyone who wants an assistant that automates recurring work, not a chat window, is the right buyer for Portal too, as long as you treat it as prepaid metered usage and check your balance in the first week. Set up the messaging gateway on day one rather than later. It is the difference between a tool you configured once and a tool you use.

Do not buy Plus expecting a flat $20 assistant. Do not start here if you want something to install and forget, because this rewards tinkering and punishes indifference. And if your work cannot tolerate an agent with terminal access on the same machine as your files, that is a reason to run it in a container or not at all.