Nous Research · Self-hostable autonomous agent
Hermes Agent: what it really costs to run
The software costs nothing. Hosting, tokens and operations do. Here are the three lines, and when they start to weigh.
Pricing checked on 3 August 2026
What it is
Hermes Agent is the self-improving agent published by **Nous Research** under the MIT licence. The repository describes an agent with a learning loop: it creates skills from experience, improves them in use, and nudges itself to persist what it learns. Memory is full-text searchable across past sessions, with model-generated summaries.
What sets it apart from coding agents is that it is not tied to a workstation. It runs from the command line but also from Telegram, Discord, Slack, WhatsApp, Signal and email. It ships a cron scheduler and can spawn parallel subagents.
It is model-agnostic: Nous Portal, OpenRouter, OpenAI or your own endpoint, switched with one command and no code change. For deployment the documentation lists local, Docker, SSH, Singularity, Modal, Daytona and Vercel Sandbox.
What it costs
- The software€0MIT licence, full source on GitHub. No tiers, no quotas, no features held back for a paid edition.
- Hosting, entryfrom $5 per monthFigure stated by the project itself: "run it on a $5 VPS". Enough for continuous personal use.
- Hosting, serverlessnear zero when idleModal, Daytona or Vercel Sandbox. The project highlights hibernation between sessions, so cost follows usage.
- Modelsper token, at your providerThis is the deciding line. No billing goes through Nous Research if you use your own keys.
- Nous Portalsubscription, rate not checkedManaged access to 300+ models under one subscription, web search and image generation included. The amount was not verified here, so it is not quoted.
Prices as published by the vendor, in dollars and excluding applicable tax. They change often, hence the verification date.
What the quotas mean
There is no quota, because there is no vendor between you and the model. The only limits you will meet are your model provider's and your machine's.
What replaces the quota is the pace of the project: more than twenty thousand commits on the main branch and changes every week. On a service you expose to a team, that means watching updates, not an install you forget.
Two limits are documented by the project itself. On Windows, antivirus software produces false positives and the recommendation is to whitelist the whole folder. On Android through Termux, the full install pulls voice dependencies that are incompatible.
The documentation publishes no minimum CPU or memory requirement. The usable marker stays the one the project gives itself, a $5 VPS, and it holds for personal use rather than for a team.
What drives the bill
Tokens dominate everything else, by a wide margin. An agent keeping searchable memory and spawning subagents re-reads a lot of context each turn. The $5 VPS will not move, the model bill will.
The learning loop has its own cost. Creating skills, improving them in use and summarising past sessions are extra calls. That is also what makes the tool interesting, so it is a line to budget rather than to switch off.
The built-in scheduler is the classic trap. A badly scoped recurring task runs at night, at weekends and through the holidays with nobody watching. A spending cap at your model provider takes five minutes to set.
The most underestimated line is human. An agent reachable from Slack and WhatsApp becomes a team service, with availability expectations. That operating time appears on no invoice and often weighs more than hosting.
Alternatives
- Claude CodeAnthropic$0 to $200 per month
A development agent, not an ops agent. Better on repository code, absent from team messaging and long-term memory.
- n8nn8n GmbHself-hosted or subscription
Deterministic flow automation. Preferable whenever rules are fixed: an agent on a predictable chain adds risk with no gain.
- Assistants managésOpenAI, Anthropicper-seat subscription
Nothing to operate, support included, and data that leaves your premises. That is the real trade-off against a self-hosted agent.
Where we stand
This block is a field opinion, not a record. Everything above it can be checked against the vendor.
Disclosure first: kengdev sells a packaged deployment of this project under the name Hermes. So this page is written by someone with an interest in your attention. Everything above comes from the repository and its documentation, and the software is free: you can check every line and install it without us.
The honest calculation runs like this. For personal use, installation takes one command and the real cost comes down to your tokens. There is no reason to pay anyone, and that is what we say when asked.
What changes at team scale is that the agent becomes a service. You have to decide what it may do alone, keep a trace of its actions, cap its spending, and follow a repository that moves every week. That work has nothing to do with installation, and it is the only part it is reasonable to pay someone for.
Which plan for which profile
- Personal useLocal or $5 VPS
One install command, your own model keys, nothing else to pay. Start there before anything else.
- Intermittent useServerless
Modal, Daytona or Vercel Sandbox hibernate between sessions, so the bill follows real use rather than an always-on server.
- Team with ops capacitySelf-hosted, by you
If someone on your side can run a self-hosted service, the project documentation is enough. Keep the budget for the action policy and the caps.
- Team without opsAssisted deployment
The only case where our offer makes sense: operations, guardrails and watching a repository that changes every week.
Frequently asked
- Is Hermes Agent free?
- Yes, entirely, under the MIT licence. There is no paid tier, no quota, no reserved feature. What costs sits elsewhere: hosting from $5 a month according to the project, and above all tokens at your model provider.
- What server do you need to run it?
- The project states a $5 VPS is enough, without publishing a minimum CPU or memory requirement. That marker holds for personal use. For a team, the constraint comes less from the machine than from the availability the service is expected to have.
- Do you have to use Nous Portal?
- No. The agent is model-agnostic and accepts OpenRouter, OpenAI or your own endpoint, switched with one command. Nous Portal bundles 300+ models under one subscription, which simplifies billing without being required.
- How does it differ from a coding agent?
- A coding agent lives in a repository and on a workstation. This one lives on a server and is driven from Slack, WhatsApp, Telegram or email, with persistent memory and a scheduler. The two complement each other more than they replace each other.
- Is it production ready?
- The software runs, the real question is what you allow it to do. An agent reachable by messaging, able to write into your tools and to trigger itself, needs a written action policy, an audit trail and a spending cap. Without those it is a prototype exposed to your team.
- Why is kengdev writing about it?
- Because we use it daily and sell its packaged deployment under the name Hermes. That is a direct interest, so we write it down. The software being free and the code public, every claim on this page can be checked without us.
Sources
Put it into practice
The fourth profile in the grid above, a team with no ops capacity, is exactly what our Hermes offer covers. The software stays free and yours.
Scope your first AI agent
20 minutes to review your tools, data and the first useful case. No jargon, no commitment.