Where does your source code actually go?
This is the question that separates these seven tools, and it has three distinct answers.
Nowhere: Tabby has no cloud component and no telemetry, running entirely on servers you control, so you remain the sole data controller. Refact.ai self-hosts the full stack on your own hardware including inference.
Wherever you configure: Continue routes inference to whatever provider you point it at — a European API, a US one, or a local Ollama or vLLM server on your own machine.
To a vendor: JetBrains AI Assistant sends context to EU or US regions depending on configuration, with enterprise no-training terms and a local-model option; Lovable processes project data in the EU but runs model inference through third-party providers. Both are reasonable; neither is "your code never leaves".
Which one fits your existing setup?
The migration cost is usually larger than the licence cost, so start there.
On JetBrains IDEs: JetBrains AI Assistant, because it needs no editor change and draws on the IDE's actual code index — resolved symbols and project structure rather than a text window. AI Pro is $10 per month, AI Ultimate $30, and it is bundled into the All Products Pack many teams already buy.
On VS Code and unwilling to move: Continue, Apache 2.0 and free, with completion, chat, agent mode and custom context providers. You pay only for whatever model you choose to call.
Building something from scratch rather than editing an existing codebase: Lovable, which produces a working React and Supabase application from a prompt, then hands you the code.
What does self-hosted AI coding actually cost you?
Hardware and operational effort, in exchange for eliminating a category of risk that some organisations cannot accept at all.
Tabby runs on consumer-grade GPUs and needs no cloud service, indexing your repositories so its answer engine responds from your actual code rather than general knowledge. Apache 2.0 with no per-seat licensing, and an Enterprise tier adding SSO and access control.
Refact.ai goes further with fine-tuning on your own codebase, so completions match your conventions rather than the average of GitHub, plus an agentic mode for multi-step edits. BSD-3-Clause, free self-hosted, $10 per month for the cloud version.
The honest trade-off is model quality. Locally runnable models trail frontier models on complex reasoning, and both projects are less polished than commercial assistants. For a bank, a defence contractor or a healthcare vendor whose code cannot leave the network, that trade is not close.
Is prompt-to-app a real category or a demo?
Real, with a specific and limited job: getting from an idea to something clickable in an afternoon.
Lovable generates a complete React and TypeScript project with Supabase for data and auth, live preview while it builds, visual edit mode alongside prompting, and custom domains. Because the output is conventional React rather than a proprietary format, and because GitHub sync is two-way, the code is genuinely yours — you can clone it and walk away, which is the opposite of most no-code lock-in.
What it does not do is produce production code unreviewed. The output needs engineering review before it carries real users, and credit costs climb with heavy iteration.
Lovable AB is Swedish with EU project-data handling, though model inference runs through third-party providers — so a strict "no US inference" requirement rules it out.
Is there a European model behind any of this?
Mistral, and it is worth understanding as infrastructure rather than as a competing assistant.
Mistral publishes open-weight models including Codestral for code, which means Continue, Refact.ai and Tabby can all be pointed at European or locally hosted Mistral inference. That is how a team gets a European model behind a familiar editor, and it is the practical reason Mistral sits at the top of this category.
Hugging Face plays the adjacent role: the Model Hub is where those open models are distributed, the Transformers library is how they are loaded, and the Inference API is one way to serve them. Pro is $9 per month; the hub itself is free.
Everywhere else in this category, the frontier models are third-party and mostly American, including inside JetBrains AI Assistant and Lovable. Saying so plainly is more useful than implying otherwise — the European part is the tooling and the jurisdiction, and with Mistral or a local model, the inference too.
