Every European AI coding assistant reviewed
Paris, France
Founded 2023
Free tier / pay-per-use API
Free tier
Best for: Teams wanting European or self-hosted model inference behind their coding tools
Mistral belongs in this category as infrastructure rather than as an assistant you install. It publishes open-weight models including Codestral for code generation, which means the tools that let you choose a backend — Continue, Refact.ai, Tabby — can all be pointed at Mistral inference running in Europe or on your own hardware. That is the practical route to a European model behind a familiar editor.
The model lineup behind it is genuinely competitive rather than a regional alternative: Mistral Large 3 is a sparse mixture-of-experts model with 41 billion active of 675 billion total parameters and a 256,000-token context window, with Magistral covering reasoning and the compact Ministral 3 family at 3B and 7B for edge and resource-constrained deployment. Several of these ship as open weights, which no US frontier lab matches at that tier.
For a development team the arithmetic is straightforward. API usage starts from roughly €1 per million input tokens for Mistral Small, and self-hosting an open-weight model costs only compute, which beats per-token pricing once volume is real. Mistral AI operates from Paris, so API calls carrying code context stay in EU jurisdiction under French law rather than falling under the CLOUD Act.
What Mistral AI does well
- Open-weight models including Codestral you can self-host
- Powers Continue, Refact.ai and Tabby as a configurable backend
- API in EU jurisdiction under French law
- From about €1 per million input tokens for Small
- Compact 3B and 7B models for local and edge use
Where Mistral AI falls short
- Not an editor plugin — you need a tool in front of it
- Frontier coding benchmarks still led by US labs
- Fast release cadence requires attention to stay current
- Self-hosting requires GPU capacity
Standout feature. The European inference layer the rest of this category can point at — and the reason "European AI coding" is a real option rather than a compromise.
Paris, France (and New York)
Free tier / Pro from $9 per month / Enterprise
Free tier
Best for: Developers and researchers working directly with open models
Hugging Face is where open machine learning is distributed, and for a development team its relevance is practical rather than philosophical: the Model Hub is how you find an open code model, the Transformers library is how you load it, and the Inference API is one way to serve it without building infrastructure.
Around that sit Datasets for training data, Spaces for interactive applications that let you try a model before committing, and deployment tooling that shortens the path from experiment to endpoint. For anyone evaluating whether a self-hosted assistant is viable, this is where that evaluation happens.
Hugging Face Inc. operates from Paris and New York, which is worth stating precisely: it is a French-rooted company with substantial US operations, not an EU-only vendor, so it does not satisfy a strict data-residency requirement on its own. What it does satisfy is the sovereignty argument in a different form — the models are downloadable, and a downloaded model has no jurisdiction at all. Free tier, Pro from $9 per month, Enterprise above that.
What Hugging Face does well
- The distribution point for open models including code models
- Transformers library and Inference API shorten the path to production
- Spaces let you evaluate a model before committing
- Downloadable models have no jurisdiction question at all
- Free tier, Pro from $9/month
Where Hugging Face falls short
- Paris and New York operations, not EU-only
- A platform, not a coding assistant you install
- Requires ML familiarity to get value from directly
- Hosted inference is not the cheapest at scale
Standout feature. The place a self-hosted coding assistant gets its model — which makes it the starting point for anyone who wants inference on their own hardware.
Prague, Czech Republic
Founded 2000
Free tier / AI Pro from $10 per month / AI Ultimate from $30 per month
Free tier, bundled in All Products Pack
Best for: Teams already standardised on IntelliJ, PyCharm, GoLand, Rider or WebStorm
JetBrains AI Assistant is the option with the lowest adoption cost for a large group of teams, because it requires no editor migration at all. It lives inside the IDEs those teams already use, offering context-aware completion, in-IDE chat, the Junie coding agent, commit message generation and multi-file edits.
Its structural advantage over editor-based rivals is the code index. JetBrains IDEs already maintain a resolved model of your project — symbols, types, references, structure — and the assistant works from that rather than from a window of text, which shows up in refactors and cross-file changes where textual context is not enough.
On data handling, JetBrains offers EU and US regions, enterprise terms excluding your code from training, and local model support through LM Studio or Ollama for work that must stay on the machine.
JetBrains s.r.o. has operated from Prague since 2000, which is a longer track record than anything else in this category. A limited free tier, AI Pro from $10 per month, AI Ultimate from $30, and inclusion in the All Products Pack many teams already buy. The Junie agent is less mature than Cursor's agent mode, and the frontier models underneath are third-party rather than European.
What JetBrains AI Assistant does well
- No editor migration for JetBrains teams
- Uses the IDE's resolved code index, not just text
- Local model option via LM Studio or Ollama
- Bundled into the All Products Pack
- Czech vendor under EU jurisdiction since 2000
Where JetBrains AI Assistant falls short
- Only works inside JetBrains IDEs
- Junie agent less mature than Cursor's agent mode
- Frontier models are third-party, not European
- AI Ultimate at $30/month is not cheap per seat
Standout feature. It reads the IDE's real code index rather than a text window — which is why its multi-file refactors land where text-based assistants guess.
Stockholm, Sweden
Founded 2023
Free tier / Pro from $25 per month / Teams from $30 per user per month
Free tier
Best for: Founders and product teams building prototypes and internal tools fast
Lovable turns a description into a working application, and what makes it more than a demo is what it produces: a conventional React and TypeScript project with Supabase handling data and authentication, not a proprietary format locked to the platform. Live preview shows the app as it builds, visual edit mode lets you adjust the interface directly instead of prompting for every change, and custom domains cover publishing.
The ownership story is the strongest argument. Two-way GitHub sync means the repository is yours throughout — clone it, hand it to your engineers, host it anywhere, stop paying Lovable. That is close to the opposite of typical no-code lock-in, where leaving means rebuilding.
Its natural place in a team is the front of the pipeline: prototypes, internal tools and the first clickable version of an idea, produced in an afternoon rather than a sprint.
The output needs engineering review before it carries real users, and credit costs climb with heavy iteration. Lovable AB is based in Stockholm with EU project-data handling, though model inference runs through third-party providers, so a strict no-US-inference requirement rules it out. Free tier, Pro $25 per month, Teams $30 per user.
What Lovable does well
- Produces conventional React and TypeScript you own
- Two-way GitHub sync keeps exit cost near zero
- Visual editing alongside prompting
- Swedish company, EU project-data handling
- Idea to clickable application in an afternoon
Where Lovable falls short
- Output needs engineering review before production
- Credit costs rise with heavy iteration
- Model inference is not European
- Not suited to editing a large existing codebase
Standout feature. Two-way GitHub sync: the generated project is a normal repository you can clone and keep, which is not how prompt-to-app usually ends.
Amsterdam, Netherlands
Founded 2022
Free and open source self-hosted / Cloud from $10 per month
Free, self-hosted
Best for: Regulated teams that cannot send source code to a third-party service
Refact.ai is a complete AI coding assistant that runs on your own hardware — completion, in-IDE chat and an autonomous agent for multi-step edits, with inference included in the self-hosted stack rather than called out to a cloud. For a bank, a defence supplier or a healthcare vendor, that is the difference between adopting AI assistance and not adopting it.
Fine-tuning on your own codebase is the feature that separates it from other self-hosted options. Completions learn your conventions, your internal libraries and your naming rather than reproducing the statistical average of public GitHub, which is where generic assistants get irritating in a large proprietary codebase. Plugins cover VS Code and JetBrains, so nobody changes editors.
It is open source under BSD-3-Clause and free when self-hosted, with a managed cloud from $10 per month for teams that want the tooling without the operations. Refact is based in Amsterdam under EU jurisdiction. The costs are real: you need GPU capacity, someone has to run it, locally runnable models trail frontier quality on complex reasoning, and the polish is below commercial assistants.
What Refact.ai does well
- Runs entirely on your own hardware including inference
- Open source under BSD-3-Clause, no per-seat cost self-hosted
- Fine-tuning on your private codebase
- Includes an agentic mode for multi-step edits
- Amsterdam-based, EU jurisdiction
Where Refact.ai falls short
- Requires GPU capacity and operational effort
- Local models trail frontier quality
- Less polished than commercial assistants
- Fine-tuning needs enough code to be worth it
Standout feature. Fine-tuning on your own codebase, self-hosted: completions that match your conventions without a line of code leaving the building.
Distributed
Founded 2023
Free and open source; you pay only your model provider
Free
Best for: Developers who want to keep VS Code but control where inference happens
Continue is a plugin rather than a product, and that is its entire value: it gives VS Code and JetBrains autocomplete, chat and agent mode while leaving the choice of model entirely to you. Point it at Mistral Codestral, at a European API endpoint, at a US provider, or at a local Ollama or vLLM server — the decision about where your code is processed becomes a configuration line rather than a vendor lock.
Context control is the second reason to choose it. Custom context providers make explicit what is sent with each request, so the assistant's inputs are inspectable rather than opaque — which is what allows a security team to actually review the deployment rather than accept a vendor statement about it.
It is Apache 2.0 and free, with no cost beyond whatever model provider you use. Two things belong in the governance note: it requires configuration and API key management rather than working out of the box, and the Continue project itself is US-incorporated even though nothing about your inference has to be. The agent mode also trails commercial rivals in maturity.
What Continue does well
- Point it at any model, including EU and fully local ones
- Apache 2.0 and free — you pay only your model provider
- Explicit, inspectable context control
- Works in both VS Code and JetBrains
- No editor migration required
Where Continue falls short
- Requires configuration and API key management
- Agent mode behind commercial rivals
- The project itself is US-incorporated
- Quality depends entirely on the model you configure
Standout feature. Where your code goes is a line of configuration rather than a vendor decision — which is the only version of this choice you can actually audit.
Distributed
Founded 2023
Free and open source; Enterprise tier for SSO and access control
Free
Best for: Organisations with a hard requirement that source code never leaves the network
Tabby is built on a single premise: no code, no telemetry and no prompt ever leaves your infrastructure. There is no cloud component to disable and no optional data sharing to configure away — self-hosting is the only deployment model, which makes the compliance conversation short in a way that no vendor assurance can match. You remain the sole data controller.
It runs on consumer-grade GPUs, which keeps the hardware requirement realistic for a team rather than a data centre project, and it indexes your repositories so its answer engine responds from your actual code rather than general knowledge. Team management, an OpenAPI interface and completion across the usual languages round it out.
Apache 2.0 with no per-seat licensing, and an Enterprise tier adding SSO and access control for larger deployments. The limitations are the honest ones for this approach: locally runnable models are weaker on complex reasoning than frontier models, agentic capability is limited compared with commercial assistants, and the maintainer TabbyML is a US company even though the software runs entirely on your hardware — which for most buyers is immaterial, since nothing is sent to them.
What Tabby does well
- No cloud component and no telemetry at all
- Runs on consumer-grade GPUs
- Answer engine grounded in your indexed repositories
- Apache 2.0, no per-seat licensing
- You remain sole data controller
Where Tabby falls short
- Local models weaker on complex reasoning
- Limited agentic capability
- Maintainer is a US company
- You operate and maintain the deployment
Standout feature. There is no cloud mode to accidentally leave enabled — self-hosting is the only way it runs.
Lisbon, Portugal
Founded 2012
Free for individual developers (Developer plan) / Team from $18 per developer/month annual or $21 monthly, up to 30 developers / Business: custom pricing; free forever for open-source projects
Best for: Teams wanting AI pull-request review and auto-fix from an EU vendor
Codacy sits earlier in the pipeline than most tools in this category: instead of writing code, it reviews and cleans up what other tools — human or AI — already wrote. AI Reviewer comments on pull requests, Guardrails auto-fixes AI-generated code before it reaches the editor, and for files with low coverage it can AI-generate the missing unit tests. Across 49 languages, that covers the housekeeping a fast-moving team using Copilot, Cursor or Continue tends to skip.
The company operates under the entity Qamine Portugal S.A. from Lisbon, founded in 2012, and states SOC 2, ISO 27001 and GDPR compliance on its own site — a longer paper trail than most newer AI review tools carry. The Developer plan is free for individuals, and open-source projects get the product free forever, which is how it gets tried before it gets bought.
Team pricing starts at $18 per developer per month annually (up to 30 developers), rising to a custom Business quote for larger, security-conscious organisations. It is not an editor plugin that writes code for you — it is the layer that checks the code afterwards, which makes it a companion to an assistant in this category rather than a substitute for one.
What Codacy does well
- AI Reviewer comments directly on pull requests
- Guardrails auto-fixes AI-written code before merge
- AI-generates missing unit tests for low-coverage files
- Free for individual developers and for open source
- Portuguese vendor stating SOC 2, ISO 27001 and GDPR
Where Codacy falls short
- Reviews code rather than writing it — needs a generator alongside it
- Business-tier pricing is quote-only
- No published on-premise option
- Team plan caps at 30 developers before Business pricing applies
Standout feature. Guardrails catches and auto-fixes AI-written code before a human ever reviews it — quality control built for an AI-assisted pipeline.
London, United Kingdom
Founded 2018
Free for open-source repos / Pro $12 per seat/month / Team $24 per seat/month / Enterprise custom pricing with self-hosting
Free trial on paid plans
Best for: Teams wanting an AI reviewer on every pull request without a big platform
Sourcery does one job: it reads every pull request and leaves the comments a senior reviewer would, including line-by-line suggestions, change summaries and diagrams that make a large diff easier to reason about. Security scans run alongside the review, and the Team plan's bring-your-own-LLM option lets a team point Sourcery at a model of its choice rather than accepting a fixed one.
Sourcery.AI Limited is a UK company incorporated in 2018, registered in London, with UK law and London courts governing its terms — a shorter, plainer jurisdiction story than tools that split EU and US regions. Both paid plans offer a free trial, and open-source repositories get Pro-level review for up to three repositories at no cost.
Pricing is transparent and modest next to the coding assistants elsewhere in this category: Pro at $12 per seat per month, Team at $24, Enterprise on request with self-hosting. What it does not do is write or complete code — it is a reviewer, not a generator, so it pairs with rather than replaces a tool like Continue or JetBrains AI Assistant.
What Sourcery does well
- Comments on every pull request automatically
- Change summaries and diagrams for large diffs
- Bring-your-own-LLM on the Team plan
- UK company, plain single-jurisdiction terms
- Transparent per-seat pricing from $12/month
Where Sourcery falls short
- Review only — does not generate or complete code
- Free tier limited to three open-source repositories
- Self-hosting is an Enterprise-only feature
- Smaller feature set than full coding-assistant platforms
Standout feature. Bring-your-own-LLM on the Team plan: the one reviewer here that lets you swap the model underneath it rather than accepting the vendor's choice.
Malmö, Sweden
Founded 2015
Free Cloud Service for open-source projects / CodeScene On Premise, Cloud and the CodeScene ACE AI add-on priced by quote — no public price list
Free trial
Best for: Engineering leads who need to prioritise technical debt by actual business impact
CodeScene answers a different question than the rest of this category: not "write this code for me" but "which part of this codebase is actually dangerous". Its behavioural code analysis combines code metrics with the history of how a file has been changed and by whom, surfacing hotspots — code that is both complex and frequently modified — as the places refactoring effort actually pays off, rather than treating every code-smell warning as equally urgent.
CodeScene AB is Swedish, founded in 2015 by Adam Tornhill in Malmö, and its Core Product is available either as downloadable software you run yourself (CodeScene On Premise) or as CodeScene Cloud. The CodeScene ACE add-on layers AI-based automated refactoring on top, explicitly built on large language models from third-party providers — an opt-in step up from the self-hostable core rather than something forced on every user.
There is no public price list: Community/open-source projects get a free Cloud Service, and On Premise, Cloud and ACE are all quoted through a sales conversation. That is a heavier sales process than most tools here, and CodeScene is an analysis and prioritisation tool rather than an autocomplete or chat assistant — it tells you where to focus, and something else still has to write the fix.
What CodeScene does well
- On Premise option keeps analysis off any vendor's servers
- Behavioural analysis prioritises technical debt by real impact, not just code smells
- CodeScene ACE adds AI-based automated refactoring as an add-on
- Swedish vendor with a decade of research behind the method
- Free Cloud Service for open-source projects
Where CodeScene falls short
- No public pricing — quote-only for every paid tier
- ACE add-on sends code to third-party LLM providers
- Analysis and prioritisation, not code generation
- Best value shows up on larger, older codebases
Standout feature. It ranks technical debt by how the code actually gets changed and by whom — a different, evidence-based answer to "what do we fix first".
London, United Kingdom
Founded 2016
Starting at $1,500 for 5,000 lines of net-new test coverage ($0.30 per line); free trial available
Free trial
Best for: Java teams that need unit test coverage without code ever leaving the building
Diffblue Cover is an autonomous agent that writes Java unit tests, and its approach is the opposite of a cloud LLM call: it uses reinforcement learning trained to produce tests that are guaranteed to compile, run and pass, operating on-prem so that, in the company's own words, code stays within your environment and is "never seen and never shared".
For a bank or defence supplier that has ruled out sending source to any external inference API, that is a materially different guarantee than an opt-out setting.
Diffblue Limited traces back to an Oxford University AI research group and early seed funding from Oxford Science Enterprises, and is registered in London (incorporated 2016). It plugs into the IDE and CI/CD pipeline rather than asking a developer to change workflow, and every generated test is verified before it counts.
Pricing runs on coverage rather than seats: from $1,500 for 5,000 lines of net-new test coverage, roughly $0.30 per line, with a free trial to evaluate on one project first. That is a real cost for a team without much legacy Java debt to cover, and the scope is narrow by design — unit tests only, Java only, not a general coding assistant.
What Diffblue does well
- Runs on-prem/in CI — code is never sent externally
- Reinforcement-learning agent, not a third-party LLM call
- Every generated test verified to compile and pass
- Free trial to evaluate on one project first
- Oxford University research origin, UK-registered
Where Diffblue falls short
- Java only — no other language support
- Test generation only, not a general coding assistant
- Pay-per-coverage pricing adds up on very large codebases
- Best suited to teams with real legacy test-debt to close
Standout feature. "Never seen and never shared": a reinforcement-learning agent that writes verified unit tests without a single line going to an external model.
Geneva, Switzerland
Founded 2008
Free tier up to 50,000 lines of code / SonarQube Team plan from about $34 per month (100k LOC) / Gitar AI Code Review from $20 per user/month annual / Enterprise on request; free forever for public OSI-licensed repos
14-day free trial, no credit card
Best for: Teams that want code-quality scanning and AI-assisted fixes from a long-established European vendor
SonarQube is the code-quality platform many of the tools elsewhere in this category get compared against, and it has added AI on top of the analysis it has run since 2008: SAST, SCA and secrets detection across 30-plus languages, plus Gitar, an AI-powered pull-request review add-on, and AI CodeFix suggestions for issues it finds.
Sonar states it is trusted by more than 7 million developers, which for a Swiss company predating most of this category by over a decade is a believable claim rather than a marketing number.
SonarSource Sàrl is based in Geneva, and the SonarQube Community Edition — the actual scanning engine — is genuinely open source under LGPLv3, published on GitHub, so a team can self-host the core analysis without paying anything or sending code to Sonar's cloud. SonarQube Server (on-premise) and SonarQube Cloud cover the commercial tiers on top of that.
Pricing is unusually concrete for this category: a free tier for private projects up to 50,000 lines of code, SonarQube Team plans from around $34 per month for up to 100,000 LOC, and Gitar AI Code Review from $20 per user per month billed annually, with a 14-day free trial and no credit card required. Public, OSI-licensed repositories get both SonarQube and Gitar free forever.
What SonarQube does well
- Community Edition open source under LGPLv3, self-hostable
- AI-powered pull-request review (Gitar) and AI CodeFix suggestions
- SAST, SCA and secrets detection across 30+ languages
- Concrete, published pricing with a free 14-day trial
- Swiss vendor with an 18-year track record
Where SonarQube falls short
- Free tier capped at 50,000 lines of private code
- AI review and auto-fix features route through third-party LLMs
- Analysis and review, not code generation
- Full enterprise feature set requires a paid plan
Standout feature. The scanning engine itself is open source under LGPLv3 — the AI layer on top is new, but the free, self-hostable core has eighteen years behind it.
