Baguette BotYou want an AI agent that browses, codes, sorts your email and runs routines, but you don't want it living on a server in Virginia. Maybe you work with client data, maybe you are a public body, maybe you simply prefer your digital life to stay under European law. Good news: in 2026 this is possible. Less good news: "hosted in France" covers very different realities.
This guide sets out the four realistic options, from the most hands-on to the most turnkey, with what each really protects. For the legal background (GDPR transfers, the CLOUD Act, SecNumCloud), see our guide to sovereign AI agents.
First, split the problem in two
Every agent has two halves, and they can live in different countries:
- The agent's computer: the machine where it runs commands, opens a browser and keeps files, sessions and memory.
- The model: the language model that reads each step and decides the next one. It runs at the model provider's.
An agent "hosted in France" whose every thought goes to a US model API keeps your files in France but sends your content to the US one prompt at a time. Both halves matter, and each option below handles them differently.
Option 1: self-host an open agent on a European server
How it works. You rent a server from a European provider (OVHcloud, Scaleway, Hetzner and others have data centres in France or the EU) and install an open agent such as Hermes Agent. You connect it to a model API, or run an open-weight model yourself.
What it protects. The computer half is entirely under your control: files, memory and logs stay on a machine you chose. If you also run an open-weight model on that server or on a European GPU cloud, nothing leaves your infrastructure.
What it costs. A small VPS costs a few euros a month. A GPU able to run a good open model at agent speed costs much more. Hermes Agent requires models with at least 64k tokens of context, which rules out the smallest local models. The real cost is your time: installation, updates, backups, monitoring, and security hardening. An agent with shell access and a Telegram gateway is an attractive target if it is badly configured.
Best for: developers and technical teams who enjoy running their own stack.
Option 2: use a European model provider
How it works. Whatever runs the agent, you point it at a model served in Europe. The best-known European model maker is Mistral AI, a French company that publishes both commercial and open-weight models and has been building its own compute capacity in France since 2025.
What it protects. The model half: your prompts are processed by a European company rather than sent to a US provider. Combined with a computer hosted in Europe, this is the closest thing to an end-to-end European agent.
What it costs. Mainly the trade-off on capability: depending on the task, the best US models may still do better on long autonomous work or complex code. For many everyday agent jobs (sorting mail, summarising, filling in documents, monitoring pages) a Mistral model is more than enough. Test on your own tasks.
Best for: anyone whose content is sensitive, and who can accept choosing models from a shorter list.
Option 3: a hosted agent platform in France
How it works. A provider runs everything for you: the agent's computer, the browser, the apps, the scheduling. You sign up and give instructions. This is what US platforms such as Grok Bot and Meta Muse do, with computers in the US. Baguette Bot does the same with computers in Paris.
What it protects. It depends entirely on the provider, so ask the questions in our checklist. In Baguette Bot's case:
- each account gets its own Linux computer hosted in Paris by a European host, with its files, sessions and encrypted secrets;
- you choose the model for each bot, and sovereign mode restricts all bots to Mistral AI models and switches off services hosted outside Europe;
- sensitive actions wait for your approval by default;
- you can export your bots to an open
.botfile and leave whenever you want.
To be precise about the limits: outside sovereign mode, a bot's messages go to the provider of the chosen model, often American, and our team administers the service remotely from outside the EU. Everything is listed in our privacy policy.
What it costs. A subscription instead of your time. Baguette Bot has a free plan, then paid plans from €29/month with the computer included.
Best for: people and small teams who want agents working 24/7 without becoming sysadmins, and who want apps on phone and desktop.
Option 4: a SecNumCloud-qualified provider
How it works. For data covered by France's SREN law (sensitive data of State administrations and their operators) or by strict sector rules, you need infrastructure qualified SecNumCloud by ANSSI, which includes immunity from non-EU laws. You then deploy your agent and your model on it, or use an AI service that is itself qualified.
What it protects. The highest level available in France: security audited by the State and legal immunity from extra-EU access.
What it costs. Significantly more, longer procurement, and a smaller choice of services. It is the right answer for regulated data and usually overkill for everything else.
Best for: public bodies, operators of vital importance, health and defence-related work.
Which option for which situation?
| Situation | Sensible choice |
|---|---|
| Developer, wants full control, comfortable with Linux | Option 1, ideally with option 2 |
| Freelancer or small business, client data, little time | Option 3 with a European model for sensitive bots |
| Content that must not leave European providers | Option 3 in sovereign mode, or options 1 + 2 |
| Public administration, data under the SREN law | Option 4 |
| Just want to try agents seriously | Option 3, free plan |
Questions to ask before you commit
- Where does the agent's computer run, and where are backups?
- Which model providers receive my prompts, and in which country?
- Which company operates the service, and is it subject to US law?
- Who can access my data, and from where?
- Can I require approval before the agent sends, pays or deletes?
- Can I export my agents in a documented format and delete everything?
An honest provider answers all of them in writing. If a vendor answers "it's sovereign" and nothing more, keep looking.
Sources
- ANSSI, SecNumCloud qualification: https://cyber.gouv.fr/secnumcloud
- Légifrance, loi n° 2024-449 (SREN), 21 May 2024: https://www.legifrance.gouv.fr/jorf/id/JORFTEXT000049563368
- Hermes Agent documentation, providers and context requirements: https://hermes-agent.nousresearch.com/docs/integrations/providers
- xAI documentation, Grok Bot security (computers hosted in the US): https://docs.x.ai/grok-bot/security
- Siècle Digital, Mistral AI's data centre financing, 31 March 2026: https://siecledigital.fr/2026/03/31/mistral-ai-leve-830-millions-pour-batir-son-datacenter-francais/