- Match the model to the job, and set up so you can switch: an API router or provider-agnostic tools.
- Prefer EU-hosted models and EU routers. Keep personal data away from US-hosted models unless a DPA is signed.
- Start automating with Make or n8n. API costs for typical small-business workflows are often less than a coffee a month.
- The cost to budget for is a coding harness used daily, plus your own time.
- Based in Luxembourg? The SME Package AI co-funds 70% of an eligible project, up to โฌ17,500. See AI funding in Luxembourg.
Section 01
Low-Code Workflow Platforms
Build AI-powered automations visually: no need to learn Python, LangChain, or any programming framework. Drag, drop, connect.
| Platform | Pricing | Self-Hosted? | Best For | EU Notes |
|---|---|---|---|---|
| n8n | Free to self-host (fair-code, Sustainable Use License), Cloud from โฌ20/mo (Starter, billed annually) | โ | Technical users who want full control; 500+ integrations | Can be self-hosted on EU servers |
| Make | Free tier, then from โฌ9/mo | โ | Visual thinkers; great UI, strong app ecosystem | ๐จ๐ฟ Czech company; EU data centres on AWS (Ireland; Frankfurt for Enterprise) |
| Pipedream | Free tier, then from ~โฌ88/mo (Startup, billed annually; ~โฌ134 billed monthly) | โ | Developers comfortable with light code | US |
| Zapier | Free tier, then from ~โฌ18/mo (Professional, billed annually; ~โฌ27 billed monthly) | โ | Beginners; simplest interface, 9,000+ apps | US only |
| Activepieces | Free (self-hosted), Cloud from ~โฌ18/mo (billed monthly) | โ | Open-source Make alternative; growing fast | Can be self-hosted on EU servers |
Section 02
API Routers & Orchestrators
A single gateway to access multiple AI models. One API key, switch between models easily, and keep your data in the EU.
| Platform | What It Does | Pricing | EU Compliance |
|---|---|---|---|
| EUrouter | EU-sovereign AI routing; OpenAI-compatible, 100+ models | Pay-per-use plus a markup; free plan, paid plans from โฌ39/mo | ๐ช๐บ Data stays in EU |
| Requesty EU | 600+ models through one EU endpoint, hosted in Frankfurt | Pay-per-use | ๐ช๐บ The EU endpoint keeps Requesty's own processing in Frankfurt; models hosted outside the EU are still called outside it |
| Euqai | EU-native router with intelligent model selection | No public price list | ๐ณ๐ฑ Hosted in Netherlands |
| OpenRouter | Very large model selection (hundreds of models) | Pay-per-use | Note: EU in-region routing on Business and Enterprise plans only |
| Melious | One API for 60+ open-weight models (text, image, audio, embeddings); OpenAI- and Anthropic-compatible | Pay-per-use per 1M tokens, net of VAT (e.g. GLM 5.3 โฌ1 in / โฌ3 out; cached input โฌ0.20); coding subscriptions from โฌ49/mo (currently discounted from โฌ99) | ๐ฉ๐ช German company (Saarbrรผcken); inference only in European data centres; DPA available; "never trained on your data" |
Section 03
Find Your AI Setup
Tell the picker what matters to you and it suggests a setup that fits. Every AI setup has three parts. The model is the trained AI itself: the part that reads your words and writes a reply. The inference provider is the company whose computers run that model, which makes it the company that sees your data. The harness is the app you actually work in: it sends your work to the model and can read files or run tools for you. One more term you will meet: a model with open weights is one its maker publishes, so anyone can host it, including a European company or your own computer.
Where something runs is judged by who runs the model and sees your data, not by who trained it, so each model card also names its maker. Prices are rough and in euros, checked against each provider's published pricing on 22 October 2026; a few are still due a second check, and all of them change fast, so follow the link before you budget on them. Dollar prices converted at the ECB rate of 21 October 2026 (1 EUR = 1.1234 USD; 1 EUR = 0.9256 CHF); vendors billing in dollars will charge the dollar amount.
Showing 16 harnesses, 60 models and 30 inference providers.
Harnesses (16)
The app you work in. It sends your work to a model, and can read files or run tools for you.
Aider (opens in a new tab)
Free coding assistant that saves every change as its own git step, so undoing one is easy.
- Processing
- Wherever the provider you connect runs the model; can also run fully on your computer
- Why it fits
- A coding harness, open source; reaches 46 matching models.
- Price
- Free; you pay your model provider directly
AnythingLLM (opens in a new tab)
Chat with your documents, with agents, using local models or almost any provider.
- Processing
- Wherever the provider you connect runs the model; can also run fully on your computer
- Why it fits
- A chat and cowork harness, open source; reaches 46 matching models.
- Price
- Free to self-host; a paid Desktop Pro tier exists
ChatGPT (opens in a new tab)
OpenAI's assistant app. Multi-step agent tasks need a paid plan; OpenAI has been moving this to ChatGPT Work.
- Processing
- No fixed location promised
- Why it fits
- A chat and cowork harness; reaches 2 matching models.
- Price
- Agent features on Plus, Pro, Business, Enterprise and Edu plans
Claude Code (opens in a new tab)
Anthropic's coding assistant. It officially supports only Claude models, which you can run through Anthropic or an EU region on Amazon Bedrock or Google. Some providers, such as Melious, OpenRouter, Ollama and LM Studio, publish their own setup for running other models in it, but Anthropic does not support that route.
- Processing
- In the EU; On your own computer; No fixed location promised; can also run fully on your computer
- Why it fits
- A coding harness; reaches 25 matching models.
- Price
- Included with a paid Claude plan, or pay per use through Anthropic or your cloud. With another provider you pay them; a model on your own computer costs nothing
Claude Cowork (opens in a new tab)
Claude works on your files and tasks from the desktop app, no terminal needed. By default the work goes through Anthropic; an organization can instead route it through its own Bedrock, Google or Foundry account (in preview).
- Processing
- No fixed location promised; In the EU
- Why it fits
- A chat and cowork harness; reaches 4 matching models.
- Price
- Included with a paid Claude plan (Pro, Max, Team, Enterprise), or pay per use through your organization's own cloud account (in preview)
Codex CLI (opens in a new tab)
OpenAI's coding assistant. Defaults to OpenAI, but you can point it at Mistral, Azure, OpenRouter or a local model, as long as the provider supports OpenAI's newer Responses API.
- Processing
- In the EU; On your own computer; No fixed location promised; can also run fully on your computer
- Why it fits
- A coding harness, open source; reaches 32 matching models.
- Price
- ChatGPT plan sign-in, or an OpenAI API key; other providers cost whatever they charge
Models (60)
The trained AI itself. Each card lists the places that run it and match your answers, EU first.
gpt-oss-120b (opens in a new tab)
Good for writing and email, research and analysis, coding, documents and scans (OCR); open weights.
Where you can use it: 7 places, all in the EU
- Infercom (opens in a new tab) Processing: In the EUPrice: โฌ0.22 / 1M input tokensChecked 22 October 2026
- IONOS AI Model Hub (opens in a new tab) Processing: In the EUPrice: ~โฌ0.15 / 1M input tokens (list $0.17)Checked 22 October 2026
- Melious (opens in a new tab) Processing: In the EUPrice: โฌ0.09 / 1M input tokens (excl. VAT, businesses only)Checked 22 October 2026
- OpenRouter (opens in a new tab) Processing: In the EUPrice: ~โฌ0.13 / 1M input tokens (EU endpoint price, served by Amazon Bedrock). Via eu.openrouter.ai, Business or Enterprise plan only; fails closed if no EU endpoint. 8% fee on credit purchasesChecked 22 October 2026
- OVHcloud AI Endpoints (opens in a new tab) Processing: In the EUPrice: โฌ0.08 / 1M input tokens (free rate-limited testing tier)Checked 22 October 2026
- Regolo.ai (Seeweb) (opens in a new tab) Processing: In the EUPrice: โฌ0.20 / 1M input tokens from 1 Nov 2026 (โฌ1.00 until 31 Oct)Checked 22 October 2026
- Scaleway (opens in a new tab) Processing: In the EUPrice: โฌ0.15 / 1M input tokens (first 1M tokens free)Checked 22 October 2026
Qwen3.5 9B (opens in a new tab)
Good for writing and email, research and analysis, documents and scans (OCR); open weights.
Where you can use it: 6 places, 4 in the EU
- IONOS AI Model Hub (opens in a new tab) Processing: In the EUPrice: ~โฌ0.10 / 1M input tokens (list $0.11)Checked 22 October 2026
- Melious (opens in a new tab) Processing: In the EUPrice: โฌ0.07 / 1M input tokens (excl. VAT, businesses only)Checked 22 October 2026
- OVHcloud AI Endpoints (opens in a new tab) Processing: In the EUPrice: โฌ0.10 / 1M input tokens (free rate-limited testing tier)Checked 22 October 2026
- Regolo.ai (Seeweb) (opens in a new tab) Processing: In the EUPrice: โฌ0.10 / 1M input tokens from 1 Nov 2026 (โฌ0.07 until 31 Oct)Checked 22 October 2026
- LM Studio (opens in a new tab) Processing: On your own computerPrice: Free to run; you supply the hardware; LM Studio itself is free for home and work useChecked 21 October 2026
- Ollama (opens in a new tab) Processing: On your own computerPrice: Free to run; you supply the hardware (tag qwen3.5:9b)Checked 21 October 2026
Qwen3.8 27B (opens in a new tab)
Good for writing and email, research and analysis, coding, documents and scans (OCR); open weights.
Where you can use it: 6 places, all in the EU
- Infomaniak AI Services (opens in a new tab) Processing: Switzerland (EU adequacy decision)Price: ~โฌ0.43 / 1M input tokens (CHF 0.40; 1M credits free for new users)Checked 22 October 2026
- IONOS AI Model Hub (opens in a new tab) Processing: In the EUPrice: ~โฌ0.40 / 1M input tokens (list $0.45)Checked 22 October 2026
- Melious (opens in a new tab) Processing: In the EUPrice: โฌ0.40 / 1M input tokens (excl. VAT, businesses only)Checked 22 October 2026
- OVHcloud AI Endpoints (opens in a new tab) Processing: In the EUPrice: โฌ0.40 / 1M input tokens (free rate-limited testing tier)Checked 22 October 2026
- Regolo.ai (Seeweb) (opens in a new tab) Processing: In the EUPrice: โฌ0.50 / 1M input tokensChecked 22 October 2026
- Scaleway (opens in a new tab) Processing: In the EUPrice: โฌ0.60 / 1M input tokens (first 1M tokens free)Checked 22 October 2026
Claude Haiku 5.5 (opens in a new tab)
Good for writing and email, documents and scans (OCR).
Where you can use it: 5 places, 3 in the EU
- Amazon Bedrock (opens in a new tab) Processing: In the EUPrice: ~โฌ0.10 / 1M input tokens. Regional EU endpoint, 10% above the global price. EU inference profile; call it from an EU Region other than Zurich or London to keep processing inside the EUChecked 22 October 2026
- Google Agent Platform (formerly Vertex AI) (opens in a new tab) Processing: In the EUPrice: ~โฌ0.10 / 1M input tokens. EU multi-region endpoint (EU member states only), 10% above the global priceChecked 22 October 2026
- Requesty (opens in a new tab) Processing: In the EUPrice: Price on provider pageChecked 21 October 2026
- Anthropic (opens in a new tab) Processing: No fixed location promisedPrice: ~โฌ0.09 / 1M input tokens, prompts up to 100k tokens (default global routing; Anthropic offers no EU inference option, only a US-only option at 1.1x)Checked 22 October 2026
- Claude (chat app) (opens in a new tab) Processing: No fixed location promisedPrice: Free plan; Pro ~โฌ17.80 / month ($20) or ~โฌ15.13 / month billed annually ($17); Max from ~โฌ89.02 / month ($100). No EU processing commitment on the pricing pageChecked 22 October 2026
Claude Sonnet 5.5 (opens in a new tab)
Good for writing and email, research and analysis, coding, documents and scans (OCR).
Where you can use it: 5 places, 3 in the EU
- Amazon Bedrock (opens in a new tab) Processing: In the EUPrice: ~โฌ1.96 / 1M input tokens. Regional EU endpoint, 10% above the global price. EU inference profile; call it from an EU Region other than Zurich or London to keep processing inside the EUChecked 22 October 2026
- Google Agent Platform (formerly Vertex AI) (opens in a new tab) Processing: In the EUPrice: ~โฌ1.96 / 1M input tokens. EU multi-region endpoint (EU member states only), 10% above the global priceChecked 22 October 2026
- Requesty (opens in a new tab) Processing: In the EUPrice: Price on provider pageChecked 21 October 2026
- Anthropic (opens in a new tab) Processing: No fixed location promisedPrice: ~โฌ1.78 / 1M input tokens (default global routing; Anthropic offers no EU inference option, only a US-only option at 1.1x)Checked 22 October 2026
- Claude (chat app) (opens in a new tab) Processing: No fixed location promisedPrice: Free plan; Pro ~โฌ17.80 / month ($20) or ~โฌ15.13 / month billed annually ($17); Max from ~โฌ89.02 / month ($100). No EU processing commitment on the pricing pageChecked 22 October 2026
Gemma 4 31B (opens in a new tab)
Good for writing and email, research and analysis, coding, documents and scans (OCR); open weights.
Where you can use it: 5 places, all in the EU
- Berget AI (opens in a new tab) Processing: In the EUPrice: โฌ0.25 / 1M input tokens (prepaid plan from โฌ25/month; โฌ5 trial credit)Checked 22 October 2026
- Infercom (opens in a new tab) Processing: In the EUPrice: โฌ0.20 / 1M input tokensChecked 22 October 2026
- Infomaniak AI Services (opens in a new tab) Processing: Switzerland (EU adequacy decision)Price: ~โฌ0.22 / 1M input tokens (CHF 0.20; 1M credits free for new users)Checked 22 October 2026
- Melious (opens in a new tab) Processing: In the EUPrice: โฌ0.10 / 1M input tokens (excl. VAT, businesses only)Checked 22 October 2026
- Regolo.ai (Seeweb) (opens in a new tab) Processing: In the EUPrice: โฌ0.40 / 1M input tokensChecked 22 October 2026
Inference providers (30)
The companies (or your own computer) that run the models, and so the ones that see your data.
Amazon Bedrock (opens in a new tab)
- Processing
- In the EU
- Why it fits
- Inference cloud: runs models for you behind an API; serves 3 matching models.
- Price
- Offers pay per use; per-model prices are on the model cards
Anthropic (opens in a new tab)
- Processing
- No fixed location promised
- Why it fits
- Inference cloud: runs models for you behind an API; serves 4 matching models.
- Price
- Offers pay per use; per-model prices are on the model cards
Berget AI (opens in a new tab)
- Processing
- In the EU
- Why it fits
- Inference cloud: runs models for you behind an API; a European company; serves 3 matching models.
- Price
- Offers free tier, subscription; per-model prices are on the model cards
ChatGPT (opens in a new tab)
- Processing
- No fixed location promised
- Why it fits
- Chat app: sign in and use it in the browser; serves 2 matching models.
- Price
- Offers free tier, subscription; per-model prices are on the model cards
Claude (chat app) (opens in a new tab)
- Processing
- No fixed location promised
- Why it fits
- Chat app: sign in and use it in the browser; serves 4 matching models.
- Price
- Offers free tier, subscription; per-model prices are on the model cards
Deepgram (opens in a new tab)
- Processing
- In the EU
- Why it fits
- Inference cloud: runs models for you behind an API; serves 1 matching model.
- Price
- Offers free tier, pay per use; per-model prices are on the model cards
This is not an exhaustive list, though we have tried to make it thorough. If something you use is missing, it may be outside the constraints of this table's informational purposes.
Details and prices checked 22 October 2026; each provider line shows its own check date.
Picking a model is usually only the first step: the real value comes from turning your documents and data into searchable knowledge systems your team can query. We build document classification and media-monitoring workflows like this on infrastructure you control for NGOs and human-rights organizations.
Run it locally
You probably already use local AI without knowing it. Running your own model is the step that gives you full privacy, and you can climb to it one rung at a time. Our post Where to start with AI if you're just getting started walks through each step.
- Already on your Mac or PC: photo recognition and transcription run locally.
- Gemma 4: a lightweight open model.
- Ollama or LM Studio: connect the app to the model.
- OpenWork: a free, open-source harness.
Section 04
Coding Harnesses
When someone says "I use Claude Code," they are naming two things at once. There is the model, the intelligence, and there is the harness, the program wrapped around it that gives the model hands: it reads your files, writes changes, runs commands, remembers what it already tried, and decides when to stop and ask you something.
This distinction is worth holding onto, because the two are separate purchases. The same model behaves very differently in two harnesses, and it is the harness, not the model, that determines where your code and prompts are sent. If sovereignty matters to you, this is the layer where you win or lose it. We wrote up how to put other models behind a harness like Claude Code in Claude Code is Dead. Long Live Claude Code.
| Harness | Open source | Models it runs | What you pay | Best for |
|---|---|---|---|---|
| Claude Code Anthropic ๐บ๐ธ | No | Claude, but on the cloud and region you pick. Other models via a gateway (see below) | Included with Claude Pro (~โฌ18/mo) or Max (from ~โฌ89/mo), or API pay-per-use | One of the most advanced harnesses for tool calling and multi-agent orchestration, but it works best with Anthropic models |
| opencode Community | Yes (MIT) | Any: 75+ providers, including Mistral and local Ollama | Free; you pay your model provider directly | Model freedom. Swap providers with a config change instead of a migration |
| Mistral Vibe Mistral AI ๐ซ๐ท | CLI yes (Apache 2.0); chat app no | Mistral (Medium 3.5, Codestral) | Free tier; Pro โฌ14.99/mo, or API usage | The sovereign default: European company, European hosting, chat and coding agent in one tool |
| Codex CLI OpenAI ๐บ๐ธ | Yes | GPT family | ChatGPT Plus/Pro sign-in, or an OpenAI API key | Sandboxed execution by default; teams already paying for ChatGPT |
| Aider Community | Yes | Any (Claude, GPT, Mistral, local Ollama) | Free; you pay your model provider directly | Git-native. Every change lands as its own descriptive commit, so undoing one thing is easy |
| Gemini CLI Google ๐บ๐ธ | Yes (Apache 2.0) | Gemini | A paid API key, or Code Assist Standard/Enterprise. Free and consumer-plan access ended on 18 June 2026; those users moved to Antigravity CLI | Very large context windows; teams already on Google Cloud |
| Copilot CLI GitHub ๐บ๐ธ | No | Multi-provider | Free tier; Pro from ~โฌ9/mo | Deep GitHub integration: pull-request aware, useful if your work already lives there |
The fully local option is Ollama behind opencode or Aider: nothing leaves your machine, no per-token cost at all. The trade is real: you need a decent GPU or a recent Mac, and small local models are noticeably weaker than frontier ones. Good for sensitive code, frustrating for hard problems.
The catch is worth understanding before you commit, because it is not obvious from the per-token price. Scaleway offers prompt caching on only a few models (DeepSeek V4 Flash, DeepSeek V4.1 Flash and Qwen3.8 27B when we checked on 20 October 2026); most of the line-up, including GLM-5.2, GLM-5.3 and Mistral Medium 3.5, has no cached-input discount. An agent resends the whole conversation (your files, its earlier reasoning, every tool result) on every single turn, and providers that cache normally charge around a tenth of the input price for the repeated part. Without that discount you pay full price for the same context, over and over, and a long agent session gets expensive far faster than the headline rate suggests.
The practical rule: excellent for one-shot and batch calls, where each request stands alone. Watch your spending carefully if you point a coding agent at it and leave it running. Providers that discount cached input, such as Melious, avoid this trap.
You are not stuck with the model your harness ships with
Worth knowing before you pick, because it undercuts the usual "Claude Code means Anthropic, full stop" assumption. There are two separate escape hatches, and they are not equally supported.
1. Same model, your choice of cloud and region. This one is official and documented. Claude Code can run against Amazon Bedrock, Google Cloud's Agent Platform (formerly Vertex AI), or Microsoft Foundry instead of Anthropic's own API: you set an environment variable and pick your region. You are still using Claude, but on Bedrock and Google's Agent Platform the inference happens in the region you chose, including EU regions, under your existing cloud contract and its data-processing terms. Foundry is the exception: Claude there runs in Global or US-only deployments, with no EU option. For a European organization that already has an EU region on Bedrock or Google, this is the straightforward way to keep an agentic coding tool inside your compliance perimeter.
2. A different model entirely, through a gateway. Claude Code talks to whatever address you put in ANTHROPIC_BASE_URL. Some providers now accept Anthropic's API format directly and publish their own Claude Code setup: Melious, OpenRouter and Requesty in the cloud, Ollama and LM Studio on your own machine. For anything else, a translation proxy (LiteLLM is the common choice) converts between Anthropic's format and whatever is behind it, such as Mistral or another open-weight model. You keep the harness you like and swap the intelligence underneath it. You pick the model with environment variables; Claude Code's built-in model menu only lists Claude models.
Choosing one, in a sentence each
- You want the least friction and you are not price-sensitive: Claude Code.
- You want to keep your options open: opencode. The provider is a line of configuration, so next year's better model is a five-minute change.
- You handle client or personal data, or sovereignty is contractual: Mistral Vibe, or opencode pointed at Mistral or Melious. If you are already committed to Claude Code, running it against an EU region on Bedrock or Google's Agent Platform gets you most of the way there (Foundry offers no EU region for Claude). Melious also speaks Anthropic's API format, so you can point Claude Code at it and run open-weight models from European data centres (route 2 above, without the proxy).
- You are nervous about an agent touching your files: Aider. Every edit is a separate git commit, so reverting is one command.
- Your whole team already lives in GitHub: Copilot CLI.
.env file. See the security questions below before any of this touches real users.
Claude Skills Worth Stealing for Any Model
A Claude Skill is a folder with a SKILL.md file (plain markdown instructions) and optional bundled scripts. The format is portable: you can drop the same instructions into opencode with Mistral, Aider, or Gemini CLI; you just lose Claude's automatic loading. Below are six worth borrowing.
| Skill | What it does | Where it's from |
|---|---|---|
| frontend-design | Builds polished UI that avoids generic "AI slop" aesthetics: strong taste, real CSS, considered typography | anthropics/skills |
| gdpr-compliance | Audits code for GDPR issues, drafts privacy notices and DPAs, answers compliance questions with article citations | orazionelson/gdpr-compliance-claude-skill |
| video-use | Edits video by conversation: transcribe, cut, colour grade, burn subtitles. No menus, no presets. | browser-use/video-use |
| systematic-debugging | Disciplined reproduce โ minimise โ hypothesise โ fix loop for hard bugs and performance regressions | obra/superpowers |
| code-review | Reviews a pull request against the plan and your coding standards: catches what self-review misses | anthropics/claude-plugins-official |
| remember | Persists session state so the next conversation picks up with full context: buffer, daily, recent, archive | thedotmack/claude-mem |
Skills can include scripts that run on your machine. Read them before you run them, and prefer skills from sources you trust.
- Clone or download the skill folder from GitHub.
- Open
SKILL.md: everything below the YAML frontmatter is plain English instructions. - Paste those instructions into your tool's system prompt, or reference the file directly: "Follow the instructions in
skills/gdpr-compliance/SKILL.mdwhen reviewing this code." - If the skill bundles scripts (Python, shell), they run on your machine regardless of which model called them.
You lose Claude's automatic "decide when to invoke" behaviour: other models will treat the skill as a suggestion rather than an imperative. Tell them explicitly when to use it.
Section 05
What Does AI Actually Cost?
Real examples with real numbers. Spoiler: it's probably less than you think.
1,000 tokens โ 750 words. A typical email is about 200โ400 tokens.Weekly Event Digest
Process 20 event descriptions, generate a categorised summary once a week
Customer Email Summarizer
Process 50 emails/day, generate daily summaries
Transcribe a 1-Hour Meeting
Whisper (self-hosted): Free
OpenAI API: ~โฌ0.32/hr
Deepgram Nova-3: ~โฌ0.23/hr
Mistral Voxtral Mini Transcribe 3: ~โฌ0.18/hr, European provider
Coding with a Harness
The one AI cost that is not loose change. A subscription is predictable; pay-per-use scales with how hard you drive it
Section 06
Low-Code Web & App Builders
Need a frontend for your AI workflow? These let you create web apps, portals, and dashboards without coding.
| Platform | Starting Price | Best For | Connects To |
|---|---|---|---|
| Softr | Free, then ~โฌ17/mo (Basic, billed annually; ~โฌ22 billed monthly) | Client portals, internal tools, simple web apps | Google Sheets, Notion, Xano |
| WeWeb | Free, then ~โฌ18/mo (Essential) | More design control; beginner-friendly | Supabase, Xano, REST APIs |
Section 07
European AI Resources & Directories
| Resource | What It Is | Link |
|---|---|---|
| European AI Atlas | Community-maintained directory of European AI companies and services | aiatlas.eu |
| AI-on-Demand Platform (AIoD) | EU-funded AI-on-demand platform | aiodp.ai |
| Mistral AI | Leading European LLM provider ๐ซ๐ท | mistral.ai |
| Infomaniak | Swiss hosting with built-in AI tools ๐จ๐ญ | infomaniak.com |
| Luxinnovation | Luxembourg's national innovation agency | luxinnovation.lu |
| SME Packages โ AI | Luxembourg State co-funding: 70% of an eligible AI project ๐ฑ๐บ | guichet.public.lu |
| Scaleway Generative APIs | Open-weight models served from French data centres ๐ซ๐ท: prompt caching on only a few models, so watch agent costs | scaleway.com |
| Ollama | Run open-weight models locally: nothing leaves your machine | ollama.com |
Section 08
Learning Resources
| Resource | What You'll Learn | Format |
|---|---|---|
| Elements of AI | The fundamentals of AI: what it is, how it works, where it's going | Free online course ๐ซ๐ฎ |
| promptingguide.ai | How to write effective prompts | Free online guide |
| learn.n8n.io | Building workflows in n8n | Free courses (sign-up required) |
| CI Newsletter | Curated AI news for business | Daily or weekly email. Sign up in the footer |
| crystallized.lu/gems | AI concepts explained simply | Blog posts |
| academy.make.com | Make platform tutorials | Free courses |
| Mistral AI Docs | Using Mistral's models and API | Documentation |
Section 09
Quick Decision Guide
Pick what you want to do, get a starting point.
"I want to automate a repetitive task" Make or n8n + an AI model
Start by picking a single task that's eating your time: sorting incoming emails, tagging leads, copying data between Google Sheets and your CRM, generating weekly reports. Don't try to automate everything at once.
In Make or n8n, you build the workflow visually: a trigger (new email, form submission, scheduled time) โ a processing step that calls an AI model to classify, summarize, or draft โ an action (update the sheet, send a Slack message, file the document). Make is friendlier for beginners and has EU data centres. n8n gives you more control and can be self-hosted on your own EU server for full data sovereignty.
Both have free tiers that are enough to get started. A typical first workflow takes 30โ60 minutes to build and saves several hours per week from day one.
Non-technical shortcut: you can ask an AI chat (Claude.ai or Mistral Vibe) to build the whole workflow for you as an importable JSON file. Describe in plain English what you want: what triggers it, what data to pull, what the AI should do with it, where the result goes, and ask for the finished n8n JSON. Then in n8n: open a new workflow โ click the โฏ menu โ Import from File (or paste the JSON directly onto the canvas). The entire workflow appears with nodes pre-wired, ready for you to plug in your credentials.
Example prompt you can copy, paste, and adapt:
I want to build a workflow in n8n that helps me keep track of what's happening in Luxembourg's networking scene, so I don't have to check the Network website every week.
Here's what I'd like it to do:
- Go to https://thenetwork.lu/events/events-calendar/ and grab the page.
- Pull out the useful bits for each event: the title, date, time, whether it's online or in person, where it is, a short description, and the link.
- Send those events to an AI model to sort and summarise them into a weekly digest aimed at busy small business owners. I'd like the events grouped into three buckets:
๐ฅ High relevance: technology, leadership
๐ Worth knowing: workshops, training, networking
๐ธ Fun: drinks and aperos- Email the digest to me via Gmail.
A few setup notes:
- I'm using n8n cloud.
- For the AI step, I'm calling Mistral through EUrouter (OpenAI-compatible API at
https://api.eurouter.ai/api/v1), with Bearer token auth. The model ismistral-medium-3.5.- I am non-technical. I need everything as set up as possible with clear instructions on how to attach my EUrouter and Gmail credentials.
- Please use as few steps as possible. Simplify where it makes sense.
- Please give me the finished workflow as a JSON file I can import straight into n8n.
"I want to summarize or draft content" Mistral Vibe (EU, free tier) or Claude.ai
Both are free chat interfaces with no setup: open the link in your browser, sign up, and start typing. Paste your text and describe what you want: "summarize in 3 bullets," "rewrite in a formal tone," "extract the action items from this meeting."
Mistral Vibe keeps your data in the EU by default, is GDPR-native, and is fastest for European languages (French, German, Spanish, Italian, etc.). Claude.ai is stronger for nuanced writing, very long documents (up to a 1M-token context window on paid plans, roughly 750,000 words), and structured reasoning tasks like reviewing contracts or analyzing spreadsheets.
The free tiers are enough for occasional use. Paid plans (about โฌ15 to โฌ18 a month) unlock the most capable models, longer documents, and higher daily limits.
"I want to transcribe meetings" Whisper (free) or Deepgram
You have three realistic routes, depending on how technical you want to get and whether the recording needs to stay private:
1. Free, on your own computer (most private): Install Whisper locally. Technical users run it as a Python command (whisper recording.mp3 --model medium.en). Non-technical users can use a friendly desktop app like MacWhisper (Mac) or Buzz (Mac/Windows/Linux): drag, drop, transcribe. Nothing ever leaves your device.
2. Free through a chat tool: Upload the audio to Mistral Vibe or Claude.ai and ask it to transcribe and summarize in one step. Fastest for one-off recordings, but the audio leaves your device.
3. Paid API (best for high volume or live): Deepgram Nova-3 is faster, more accurate with accents, and supports real-time captions. Around โฌ0.23 an hour for recorded audio: call it from Make or n8n in an automated pipeline.
"I want to read & process documents" Mistral OCR or Google Document AI
For one-off scans or simple PDFs: upload the file to Mistral Vibe or Claude.ai and ask for the extracted text, tables, or specific fields. Free, instant, no setup.
For automated pipelines (invoices arriving daily, an archive of documents to digitize, contracts to process): call the Mistral OCR API from a Make or n8n workflow. It handles complex layouts, tables, and multi-language documents in one pass. Mistral stores your data in the EU by default.
For structured field extraction (invoice number, VAT, dates, line items into a database): Google Document AI has pre-trained "processors" for invoices, receipts, IDs, and forms. Azure Document Intelligence is the comparable Microsoft offering: strongest for forms.
"I want to build a client portal" Softr with built-in tables
Softr lets you ship a login-protected client portal in an afternoon. Use their built-in database (no separate tool needed) or connect an existing Google Sheet, Airtable, or Notion workspace. Out of the box you get user authentication, per-client content, file uploads, forms, and a clean mobile-friendly design.
The free tier covers small portals. The Basic plan (~โฌ17/month, billed annually) adds a custom domain; higher plans add more users and capacity. If you want more design control or custom interactions, WeWeb is the next step up: still no-code, but closer to a full web framework.
Typical starter build: client login โ dashboard with their projects/invoices โ file upload area โ contact form. One afternoon, zero code.
"I want to stay GDPR-compliant" EUrouter or Requesty EU + Mistral
Two practical rules: keep the data inside the EU, and prefer EU-hosted models.
Route your API calls through EUrouter or Requesty EU instead of OpenRouter. Both do their own processing inside the EU (Requesty in Frankfurt, EUrouter in its EU region). The router is only half of it: a model hosted outside the EU is still called outside the EU, so pair the router with an EU-hosted model.
Pick Mistral models as your default. Mistral is a French company that stores your data in the EU by default, and its models are GDPR-native and competitive on quality. For API calls, use Mistral's EU endpoint if processing must stay in the EU too: it costs 10% more. Reserve US-hosted models (Claude, GPT) for tasks that don't involve personal data, or only after you have a signed DPA in place.
Document it. Add your AI tools to your processing register (GDPR Art. 30), note where each provider hosts data, and include them in your privacy policy. For platforms like Make, choose the EU data region at setup: it's a one-click decision that's much harder to reverse later. Even then, Make says some of its built-in AI features are still processed outside the EU, so check before you use them on personal data.
"I want to connect multiple AI models" Use an API router (see Section 02)
Instead of juggling separate accounts, API keys, and monthly bills with OpenAI, Anthropic, Mistral, Google, and the rest, use one router. You get a single API key, one dashboard, one invoice, and can swap between models by changing a single string in your workflow or code.
OpenRouter has a very large selection (hundreds of models) but is US-based, and its EU in-region routing is only on Business and Enterprise plans. EUrouter and Requesty EU are the EU-sovereign alternatives: the routing itself happens in Europe, and with an EU-hosted model so does the rest.
All can be used pay-per-use with no minimums: you pay for the tokens you use. EUrouter adds a markup on top and offers optional monthly plans. For a typical small-business workload, that's usually a few euros per month.
Section 10
The Fourteen Layers of Building a Real App
Vibe coding gets you to a prototype quickly. Shipping it to real users is the iceberg under the waterline. Here are the fourteen layers that exist whether you think about them or not: what each is, when it starts mattering, and a starter-friendly tool or pattern for each.
1. Frontend foundations
What users see and click: the HTML, CSS, and JavaScript running in their browser.
When it matters: always.
Starter pattern: React, Astro (what this site is built with), or plain HTML for a one-pager. Pick a UI kit; don't reinvent buttons.
2. APIs and backend logic
The server code that handles requests from the frontend.
When it matters: as soon as your app does anything beyond static content.
Starter pattern: Express on Node, FastAPI on Python, or Supabase to skip writing a server.
3. Database and storage
Where data persists between sessions.
When it matters: as soon as you have users with accounts, saved content, or anything that survives a refresh.
Starter pattern: Postgres. Supabase or Neon offer Postgres as a service; both have EU regions, so pick one when you create the project.
4. Auth and permissions
How users log in. Who can do what.
When it matters: the moment you have more than one user.
Starter pattern: Clerk, Auth0 (both US), or Supabase Auth. EU or self-hosted: Keycloak (open source, self-hosted) or Zitadel (Swiss, open source). Don't roll your own.
If you must roll your own: hash passwords with bcrypt or argon2: never plain text, never MD5/SHA. You'll also need secure session cookies (HttpOnly, Secure, SameSite), CSRF protection, a password-reset flow with single-use tokens, login rate limiting, and ideally TOTP for 2FA (otplib). That's the minimum, and why most people use Clerk.
5. Security and row-level security
Making sure user A can never read or modify user B's data, even if there's a bug elsewhere.
When it matters: the moment two users share an app.
Starter pattern: Postgres Row-Level Security (RLS) policies. Built into Supabase.
6. Rate limiting
Stopping abusive traffic: bots, brute-force login attempts, runaway scripts.
When it matters: before your first public deploy.
Starter pattern: Cloudflare's free tier, or express-rate-limit on Express. Self-hosted: rate limiting in your reverse proxy (Caddy or nginx).
7. Caching and CDN
Making your app feel fast for users far from your server.
When it matters: once you have international users or static assets.
Starter pattern: Cloudflare or BunnyCDN in front of your app.
8. Hosting and deployment
Where your code actually runs in production.
When it matters: the day you want anyone to use your app.
Starter pattern: Scaleway, Hetzner, or Fly.io for a real server. Vercel or Netlify (both US) for static frontends.
9. Cloud and compute
The underlying machine: CPU, RAM, disk.
When it matters: when you outgrow shared hosting or want predictable performance.
Starter pattern: a single Scaleway or Hetzner VM is enough for years for most small SaaS.
10. CI/CD and version control
Git for tracking changes. CI for running tests automatically. CD for deploying when tests pass.
When it matters: from day one for git; CI/CD when you have anyone else on the team or any tests worth running.
Starter pattern: GitHub + GitHub Actions. Sovereign alternative: self-hosted GitLab on a Scaleway or Hetzner VM, with GitLab CI built in.
11. Load balancing and scaling
Spreading traffic across multiple servers when one isn't enough.
When it matters: when you genuinely have too much traffic for one box. Most small apps never reach this.
Starter pattern: wait until you need it, then add a managed load balancer.
12. Error tracking and logs
Knowing your app is broken before users tell you.
When it matters: before your first public deploy.
Starter pattern: Bugsink (self-hosted, Sentry-compatible) for errors, or Sentry's SaaS if you don't want to host. Better Stack (Logs) for logs.
13. Availability and recovery
Backups. Restore procedures. What happens when the server dies.
When it matters: before you have a single paying customer.
Starter pattern: daily Postgres backups to Scaleway Object Storage or Infomaniak Swiss Backup. Test restoring them once a quarter.
14. Task queues and background jobs
A to-do list for your server, so you never miss anything. Slow or failure-prone work (sending an email, calling an AI model, processing an upload, handling a webhook) goes on a queue instead of running while the user waits. If a step fails, the job is retried instead of silently lost.
When it matters: as soon as your app sends emails, calls outside services, or does anything that takes more than a second or two.
Starter pattern: a queue that lives in the Postgres you already have: pg-boss on Node, Procrastinate on Python. Move to BullMQ with Redis when volume grows. All open source, running on your own server. Make sure failed jobs show up in your error tracking (layer 12).
Section 11
Six Security Questions Before You Take On Live Users
You need to be able to answer these six questions about your own product.
1. Where does the data live? EU vs. US
Picking an EU host (Scaleway, Hetzner, OVH) keeps you out of cross-border-transfer trouble. See Hosting for starter options.
2. What do you actually store? Less collected = less to leak
Audit your own forms. Don't collect a birthday you'll never use.
3. Who can see it? Authentication and access control
One sentence each on a checklist: who is logged in, what role they have, what records that role can touch. Row-Level Security in Postgres makes this enforceable, not just policy.
4. Can users get their data out, or delete it? Access, rectification, erasure, portability
GDPR in four verbs. Build a "download my data" and "delete my account" path before launch; they're easier to add early than retrofit.
5. What do you tell people you're doing? A privacy notice in plain English
Before any cookies are fired (which should only happen after the cookie banner is accepted). Name your sub-processors (hosting, AI APIs, email tools). Plain language beats legalese every time.
6. Are your secrets kept out of the AI? Keep .env away from your coding tool
Don't share your .env file with your AI coding tool. Keep real keys out of any file your AI tool can read: use placeholder values in development and store production secrets in your host's secret manager.
Section 12
Two Random Shortcuts to MVP
These are two spots where we see people run into issues: they want to connect to social media, or they want to create an app. Both can mean long negotiations with big tech companies like Meta or Apple. Here are two quick time savers. If you are running everything solo, they are often the fastest way in: see how AI orchestration works for a team of one.
PostProxy
One API connects you to 12 platforms: Instagram, Facebook, Threads, TikTok, X, YouTube, LinkedIn, Pinterest, Bluesky, Telegram, Google Business and WhatsApp. No per-platform OAuth, no quirks. You pay them; they handle the messy parts.
Progressive Web Apps (PWAs)
A web app that installs to a phone's home screen from the browser. Feels native. No app store, no native code, no store review. One codebase. For most small-business products, this is enough to launch.