Tabby
Self-hosted coding assistant designed so that no code, telemetry or prompt ever leaves your infrastructure
Quick Overview
| Company | TabbyML (open source project) |
|---|---|
| Category | AI Coding Assistants |
| Headquarters | Distributed, Open Source |
| Founded | 2023 |
| EU Presence | Runs wherever you deploy it, including EU-only infrastructure |
| Data Location | Entirely on your own servers |
| Open Source | Yes (Apache 2.0) |
| GDPR Compliant | Yes - you remain the sole processor |
| Self-Hosting | Yes - the only deployment model |
| Main Features | Self-hosted completion, repository context indexing, answer engine, team management, OpenAPI interface, consumer-GPU support |
| Pricing | Free and open source; Enterprise tier for SSO and access control |
| Best For | Organisations with a hard requirement that source code never leaves the network |
| Replaces | GitHub Copilot, Tabnine, Codeium |
Detailed Review
Tabby starts from the assumption that the answer to "where does our code go" has to be "nowhere". There is no hosted tier to accidentally enable and no telemetry to disable — the product is a self-contained server you run yourself, with IDE plugins that talk only to it.
It is Apache-licensed and deliberately modest in its hardware requirements: Tabby runs usefully on consumer GPUs, which makes an on-premise deployment realistic for a small team rather than only for organisations with a GPU cluster.
Deployment and Hardware
A single Docker command brings up the server with a chosen model. Tabby supports CUDA, ROCm and Apple Metal, so an existing workstation GPU or even a well-specified Mac can serve a small team. Larger deployments scale by adding capacity rather than by buying seats.
Because it runs on infrastructure you already own, the cost model is entirely different from per-developer subscriptions — an important point when a fifty-developer Copilot bill is being compared against a one-off GPU.
Repository Context and the Answer Engine
Tabby indexes your repositories and documentation so completions and answers reflect your own code rather than generic patterns. The answer engine lets developers ask questions about the codebase — where a function is used, how a service is configured — with responses grounded in the indexed source.
For onboarding new developers into a large internal codebase, that grounded search is often more valuable day to day than the completion itself.
Governance
The project is maintained by a US-incorporated company, but because Tabby is Apache-licensed and has no cloud component, running it creates no data-processing relationship with anyone. From a GDPR perspective you remain the sole controller and processor, which is the cleanest possible position.
Limitations to Consider
Self-hosted models are meaningfully weaker than frontier models on complex, multi-step reasoning, and Tabby's agentic capabilities are limited compared with Cursor or Claude Code. You take on the operational work — GPU capacity, container updates, model upgrades. And while the deployment is yours, the project maintainer is a US company, so this is a self-hosting choice rather than a European-vendor choice.
Who Should Use Tabby
Defence suppliers, financial institutions, healthcare software teams and anyone under contractual confidentiality obligations that rule out third-party AI services entirely. Where the requirement is absolute rather than preferential, Tabby is the most straightforward option in the category.
Pros and Cons
Pros
- No cloud component and no telemetry at all
- Runs on consumer-grade GPUs
- Repository-grounded answer engine
- Apache 2.0, no per-seat licensing
- You remain sole data controller
Cons
- Local models weaker on complex reasoning
- Limited agentic capability
- Maintainer is a US company
Alternatives to Tabby
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
Czech Republic · AI built into the IDEs your team already uses
Lovable
Sweden · Prompt-to-app builder that ships full-stack projects
See all European ai coding assistants or compare them directly against GitHub Copilot on our alternatives page.
Frequently Asked Questions
No. Tabby is self-hosted by design — there is no hosted tier, which means there is no configuration mistake that could send your code to a third party.
It supports CUDA, ROCm and Apple Metal, and runs usefully on consumer GPUs. A single workstation-class GPU is enough to serve a small team.
Both are open source and self-hosted. Refact.ai adds fine-tuning on your own codebase and a more developed agent; Tabby is simpler to deploy and has a stronger repository-grounded answer engine. Either satisfies a no-code-leaves-the-network requirement.
Running it self-hosted, you remain the sole controller and processor of the data involved, which is the simplest possible GDPR position — there is no third-party processor to assess at all.