Refact.ai
Open-source AI coding agent you host yourself, so completion, refactoring and agentic edits happen without any source code leaving your network
Quick Overview
| Company | Refact (Small Magellanic Cloud AI) |
|---|---|
| Category | AI Coding Assistants |
| Headquarters | Amsterdam, Netherlands |
| Founded | 2022 |
| EU Presence | Yes - United Kingdom (terms under English law) |
| Data Location | Local engine on your own machine; models via local runtimes or your own provider keys (Refact Cloud retired) |
| Open Source | Yes (BSD-3-Clause) |
| GDPR Compliant | Yes |
| Self-Hosting | Yes - full stack on your own hardware |
| Main Features | Code completion, in-IDE chat, autonomous agent, self-hosted inference, fine-tuning on your codebase, VS Code and JetBrains plugins |
| Pricing | Free and open source; paid Refact Cloud is being retired |
| Best For | Regulated teams that cannot send source code to a third-party service |
| Replaces | GitHub Copilot, Cursor, Tabnine |
Detailed Review
Refact.ai answers the AI coding question differently from Cursor and Copilot: rather than promising not to misuse your code, it removes the transfer entirely. The whole stack (local engine, plugins, agent, and a local model runtime if you choose one) runs on hardware you control, so proprietary source never crosses your network boundary.
The project is open source under a BSD-3-Clause licence, published by Small Magellanic Cloud AI Ltd, whose terms are governed by the laws of England and Wales. With Refact Cloud being retired in 2026, the project continues as open source and community maintained, with active development moved to a new GitHub repository.
Self-Hosted Inference
Refact runs as a local engine behind the VS Code or JetBrains plugin. You point it at a local runtime such as Ollama, LM Studio or vLLM on your own GPU, or at a provider key of your choice, and completion works exactly as it would against a cloud service. For organisations with an existing GPU allocation, the marginal cost per developer is close to zero.
The trade-off is honest: models you can host on a single GPU do not match frontier models on complex reasoning. For completion, boilerplate, test scaffolding and routine refactoring — which is the bulk of day-to-day assistance — the difference is much smaller than benchmark comparisons suggest.
Fine-Tuning on Your Own Codebase
Because inference is local, Refact can fine-tune on your repositories without that data leaving your infrastructure. For a large internal codebase with house conventions and private libraries, a fine-tuned smaller model often produces more useful completions than a bigger general model that has never seen your code.
The Agent
Refact includes an autonomous agent that can work across multiple files, run commands and iterate on failures. It is less polished than the commercial agents, but it runs under the same self-hosted guarantee, which for some organisations is the only way an agent gets approved at all.
Limitations to Consider
Self-hosting is real operational work: you need GPU capacity, someone to run the model runtime and a plan for model updates. Locally hostable models trail the frontier on hard reasoning tasks. The plugin ecosystem covers VS Code and JetBrains but not much else, and the product's polish is behind Cursor's — this is a tool chosen for its deployment model rather than for its interface.
Who Should Use Refact.ai
Banks, health organisations, defence suppliers, public bodies and any team with contractual confidentiality obligations that make sending source to a third-party AI service impossible. If self-hosting is a requirement rather than a preference, Refact.ai and Tabby are the two serious options.
Pros and Cons
Pros
- Runs entirely on your own hardware
- Open source under BSD-3-Clause
- Fine-tuning on private codebases
- Includes an agentic mode
- No per-seat cost when self-hosted
Cons
- Requires GPU capacity and operational effort
- Local models trail frontier quality
- Less polished than commercial assistants
Alternatives to Refact.ai
Other European ai coding assistants worth comparing before you decide:
Mistral AI
France · European frontier models with Codestral for code generation
Hugging Face
France · Open model hub with code models you can run yourself
JetBrains AI Assistant
Netherlands · AI built into the IDEs your team already uses
See all European ai coding assistants or compare them directly against GitHub Copilot on our alternatives page.
Frequently Asked Questions
Not when you use a local model. Refact runs as a local engine next to your IDE and sends requests only to the providers or local runtimes you configure, such as Ollama, LM Studio or vLLM on your own GPU. The hosted Refact Cloud is being retired.
A single reasonably modern GPU is enough for a small team; capacity requirements scale with concurrent developers and model size. Refact publishes deployment guidance for the supported models.
Yes, and because inference is local the training data never leaves your infrastructure. For codebases with strong internal conventions this usually beats a larger general-purpose model.
Refact is free and open source under BSD-3-Clause. The paid Refact Cloud service is being retired in 2026, so there is no paid tier any more; you pay only for any model provider you connect.
Refact.ai is built as a European alternative to GitHub Copilot: Open-source AI coding agent you host yourself, so completion, refactoring and agentic edits happen without any source code leaving your network. It will not be a like-for-like feature match in every respect, so check the review above for where the two genuinely differ before switching.
Refact.ai is one of several European AI Coding Assistants tools covered on this site. See the full AI Coding Assistants category page for how it ranks against its closest European competitors.