An open-source project launched by Nous Research and developed with its community, to which Diadems contributes.
Guide · Hermes Agent
Hermes Agent, the open-source AI agent from Nous Research
What it can do, what it takes to run on your own machine, and how Diadems runs it in France for a team.

Hermes Agent is an open-source AI agent released by Nous Research in February 2026 under the MIT licence. It gives the model of your choice tools (terminal, browser, files, web), a memory, skills it writes itself, scheduled tasks and access to messaging apps. You can install it on your own server, or hand it to Diadems, which runs it in France, in an isolated virtual machine for each agent.
What is Hermes Agent?
Hermes Agent is free software, released in February 2026 by the American lab Nous Research under the MIT licence. It is developed in the open, with a large community of contributors, Diadems among them.
A language model (GPT, Claude, Mistral…) produces text. An agent is the program around it that gives it hands: it sets the model a task, lets it choose an action (run a command, open a page, read a file), carries that action out, hands back the result, and repeats until the job is done. Hermes Agent is that program.
- Tools. Terminal, code execution, files, browser, web search, vision, voice. Commands run on the machine itself or in a separate environment: a container, a remote server.
- A memory. Two short note files that the agent keeps itself and rereads in every conversation, a search across its past exchanges, and plug-ins for a richer memory.
- Skills. When it succeeds at a task worth repeating, it writes the procedure into a
SKILL.mdfile, then corrects it with use. Nous Research calls this the “learning loop”. - Scheduled tasks. A job starts at the set time, with nobody at the screen. The result arrives on the messaging app of your choice.
- Messaging apps. Telegram, Discord, Slack, Signal, email and some twenty others. By default, it ignores anyone who has not been authorised.
- Sub-agents. A long task can be split up: isolated sub-agents work in parallel, then the agent brings their results together.
- The model of your choice. No provider is imposed: OpenRouter, OpenAI, Nous Portal, any compatible service, or a locally installed model.
- Safeguards. Commands judged dangerous can wait for approval. A blocklist always stops destructive commands.
The project moves fast: six releases in under a month at the end of summer 2026. That is its strength, and the main constraint for anyone who runs it.
What you can do with Hermes Agent
Four set-ups you can run today, showing what Hermes contributes and what Diadems adds. Volumes and schedules are indicative.

Construction, B2B services
Taking apart a tender pack
You upload the tender documents: 30 to 80 PDFs. Each document goes to a sub-agent. The agent assembles a brief: documents to supply, award criteria, penalties, dates, questions to put to the buyer. Every requirement cites its document and clause.
- Hermes provides
- Sub-agents in parallel, reading the files.
- Diadems adds
- File uploads of up to 1 GiB, step-by-step progress, reminders in Google Calendar once you agree.
- Limit
- Your team reviews the brief and decides whether to bid.
IT, software publishers, e-commerce
A nightly round of your servers
Every night, the agent connects to your servers over SSH: disk, certificates, backups, services, security patches. At 6 am, the report arrives on Telegram, alerts first, with proposed fixes.
- Hermes provides
- The scheduler, the terminal, Telegram.
- Diadems adds
- The Vault for the SSH key, kept out of the instructions.
- Limits
- An account with restricted rights. No fix without your approval. No access to servers on a private network.
Maintenance, after-sales service
On-site diagnosis from a photo
The technician photographs the rating plate or the error code. The agent identifies the machine, finds its documentation and the faults already seen, then sends back the likely procedure with part numbers.
- Hermes provides
- Vision, web search, memory.
- Diadems adds
- The Android app and a document base shared by the team.
- Limit
- A diagnostic aid. Electrical and gas safety stays with the qualified technician.
Any small business
Working directly on a Windows PC
During a conversation, you hand your PC over to the agent: renaming 400 files, auditing the machine, setting up a printer, re-keying data into software that has no API.
- Diadems only
- It goes through the Diadems Windows app, not through Hermes alone. Access covers one conversation and one machine; a banner shows that the agent is working, with a button to stop it.
- Limits
- Never at night or from a scheduled task. Access is complete under your Windows account: only grant it knowing what that means.

More examples, sorted by function: AI agent use cases in business.
Hosting Hermes Agent yourself: what it takes
Hermes Agent installs with a single command, on Linux, macOS or Windows (WSL2). Installation is not the hard part. Running it properly for a business, week after week, is harder. Eight things to plan for.
- A machine that is always on. An agent that works overnight does not live on a laptop you close. The documentation recommends 2 to 4 GB of memory and two cores; the automated browser and an advanced memory need considerably more.
- Isolation. The agent runs commands. On a shared server, one mistake affects everything else. Plan a container, a user without administrator rights or a dedicated server.
- Secrets. Model keys, messaging tokens, access to your tools: in a protected file, never in an instruction or a log.
- Access. Who can talk to the agent on Telegram or Slack? You need explicit allowlists, never an opening to everyone.
- Updates. One command is enough. At the project's pace, though, you need to read the release notes, test on a copy and know how to roll back.
- Backups. Hermes can back itself up. You still have to schedule it, encrypt the copies that hold secrets, store them elsewhere and test a restore.
- Monitoring. Full disk, a service down, a model quota used up: someone has to know before your users do.
- The team. Hermes is designed first for one person. Sharing it between colleagues, with different rights and separate conversations, takes an organisation it does not provide out of the box.
If you have these skills in-house, self-hosting is a rational choice: it is the cheapest option, and you keep control of everything.
Hosted, managed Hermes Agent in France with Diadems
Diadems installs, monitors, backs up and updates Hermes Agent for businesses. Each agent is a complete instance, with its own memory, files and access. You use it from the web portal, the Windows or Android app, and your messaging apps.
What Diadems adds
- One virtual machine per agent. Each agent runs in its own lightweight virtual machine (Kata Containers), on a dedicated network, at Scaleway, in the Paris region. A firewall shuts it off from private networks.
- Tested releases. Each official release goes through automated tests, then a pilot agent. It is applied while the agent is idle; an agent never updates itself.
- Over a hundred models. The Diadems catalogue, paid from your AI credit, shows for each model whether it is processed in France, in the EU or with no European guarantee. You can also connect your own API key or your ChatGPT or Claude subscription.
- One agent for the team. Everyone has their own conversations. Rights are set per person and per section. The agent's memory is shared.
- Built-in memory. The agent keeps facts and preferences in a memory bank of its own (Hindsight), without leaving its virtual machine.
- Its own email address. It only corresponds with the addresses and domains on its allowlist.
- Protected access. Two-factor authentication is mandatory for every member. The secrets in the Vault stay in the agent's space, with no copy in the Diadems database.
- Daily backups. With an encrypted off-site copy in the European Union.
- Support in French. Through the Diadems chat: the team takes over from the AI assistant during office hours.

What Diadems does not change
Hermes stays Hermes: its loop, its tools, its skills, its scheduled tasks, its sub-agents. When a capability is missing, Diadems does not rewrite the engine. It adds a minimal, tested patch, published on its public fork DiademsTech/hermes-agent, whose main branch stays identical to the official project. Whatever is useful to everyone is offered upstream; each patch is dropped as soon as an official release provides the equivalent.
Before you choose
- Your agent, your name. The word “Hermes” does not appear in the portal: you talk to your agent, named however you like.
- A version chosen for you. Diadems applies the one it has tested. If you want the very latest release on the day it ships, self-hosting suits you better.
- A selection of messaging apps. Telegram, Slack, Discord, Mattermost, Matrix and WhatsApp Business (Meta's official API) connect from the portal.
- Approvals to switch on. Hermes's native approvals, which hold sensitive commands until you agree, are off by default. An owner or an administrator switches them on.
- Dearer than a VPS. The price pays for isolation, tested releases, backups, team features and support. The details: pricing.
Where the data lives, and what depends on the model: an AI agent hosted in France, in the EU.
Where to run Hermes Agent
Three options: your own server, an instance rented from a hosting provider (Hostinger, Hermes Cloud, DINAO and others), or an agent managed end to end, such as Diadems. Some agencies also install it on your own infrastructure, for a quoted price.
Our view. For a single person with technical skills, a VPS or a rented instance costs far less. Hermes Cloud is the developer's own offering, with a region in Paris. Diadems is for the business that wants an agent shared by a team, hosted in France, without administering a server.
Prices, models, hosting location and what is left for you to do, offer by offer: Hermes Agent hosting, the offers compared.
Tested version and update policy
Diadems currently runs Hermes Agent v0.23.0, built from the official tag with the public patches described above, in line with the MIT licence. A new release always follows the same path.
- Nothing automatic. The project publishes a release. Diadems is notified; nothing installs itself.
- Tests. A candidate image is built on the public fork, then tested: conversations, memory, scheduled tasks, message queues, backups.
- A pilot agent. It runs on a pilot agent first. Once approved, it replaces the previous one.
- While idle. Each agent is updated when it is not working. No conversation is cut off; files, memory and settings are kept.
Frequently asked questions about Hermes Agent
What is Hermes Agent?
An open-source AI agent launched by Nous Research and developed with a large community of contributors. It is not a model: it gives the language model of your choice tools (terminal, browser, files, web search), a memory, skills it writes itself, scheduled tasks and access to messaging apps. It carries out multi-step tasks instead of only replying.
Is Hermes Agent free?
The software is free. What you do pay for is what runs it: the language models, billed by usage by their provider, and the machine that hosts it. A hosted offer adds the price of the service. At Diadems, the Solo plan costs €29 excl. VAT per month; models are paid by usage, at the provider's price, or through your ChatGPT or Claude subscription.
Is Hermes Agent open source? Under which licence?
Yes, under the MIT licence. You can use, modify and redistribute it, including commercially, provided you keep the copyright notice and the licence text.
How is it different from OpenClaw or ChatGPT?
ChatGPT is an assistant hosted by OpenAI, which you talk to in its own app. Hermes Agent is software that you install or have hosted: it acts with its own tools, keeps its memory and works without you, overnight for instance. OpenClaw is another open-source agent with similar uses. What Nous Research highlights about Hermes is its learning loop: it creates and improves its own skills.
Do you need a server to use Hermes Agent?
Not to try it: it installs on a personal computer. For it to run scheduled tasks and answer on your messaging apps at any hour, it needs a machine that is always on: a server, a VPS, or a hosted offer such as Diadems.
Which models can you use with Hermes Agent?
Whichever you like: Claude, GPT, Gemini, Mistral, DeepSeek, Qwen or a locally installed model. Hermes Agent is not a model itself and works with any compatible provider. At Diadems, you choose from the catalogue, conversation by conversation, and the memory model is set separately.
Does Hermes Agent work in languages other than English?
Yes. The agent replies in the language of your messages, with the quality of the chosen model. Its official documentation and commands are in English. At Diadems, the customer area is in English or French; the documentation is in English too, the support team answers in French for now, and the agent itself works in your language.
Where does my data go with Hermes Agent?
Self-hosted, it stays on your machine, apart from what you send to the model provider with each request. At Diadems, the agent, its files and its memory are hosted in France; AI processing depends on the chosen model, and the catalogue shows which models are processed in France or in the European Union. The tools and third-party services the agent calls have their own locations.
Sources: the official Hermes Agent GitHub repository and documentation (Nous Research); the Hermes 4 models technical report (arXiv 2508.18255); the public pages of the offers cited; the DiademsTech/hermes-agent public fork.
Updated on .
Give it its first task.
Your agent is ready in a few minutes. Or tell us what you need first: we will help you choose its first assignment.