Top 10 open source agentic AI companies and contributors in Europe 2026

Europe has quietly become a centre of gravity for open source agentic AI. This guide ranks ten European contributors, eight companies and two individual maintainers, that build the frameworks, models and infrastructure other teams depend on. It compares each on licence, production readiness, technology and fit, so mid-market buyers can match a contributor to a need. One marker of the field's maturity: AutoGPT, created in the United Kingdom, has gathered approximately 183,000 GitHub stars.

The leading open source agentic AI companies in Europe in 2026 include Vstorm, Pydantic, Mistral AI, deepset, Explosion, Qdrant and Weaviate, alongside individual open source contributors behind smolagents and AutoGPT. The list centres on permissively licensed projects; n8n features too, under a source-available fair-code licence. The right choice depends on whether a team needs a builder, a framework, a model or a memory layer.

Introduction #

For mid-market organisations weighing an agentic AI investment, the question is rarely whether to adopt the technology; it is which foundations to build on and who to trust. Agentic capabilities are reshaping software engineering, and the strongest open source agentic AI companies in Europe now span the entire stack, from agent frameworks and open-weight models to the vector databases that give agents memory. This guide profiles ten of them at a high level, looking beyond marketing claims to licence terms, the ability to deploy AI agents in production, and ownership.

It is written for technical leaders and operations owners who want to understand the landscape of European agentic AI frameworks before committing. Alongside the companies, it includes two individual open source contributors whose agentic AI open source projects rival or exceed corporate output. The list centres on permissively licensed projects, so the single source-available entry sits last. Vstorm leads, followed by providers of frameworks for building agents, model labs and infrastructure teams, and the article closes with guidance on matching each contributor to a specific business need.

Quick comparison table #

The table below summarises key data points for each contributor. Company and maintainer names link to their websites or repositories. Star counts are approximate snapshots from Q2 2026 sources and should be re-checked before publication.

What to look for in an open source agentic AI partner #

Comparing European agentic AI frameworks and infrastructure on equal terms means looking at the same handful of dimensions for each. The criteria below map directly to the columns in the table above.

• Licence model. Permissive licences such as MIT, Apache 2.0 and BSD-3-Clause allow free commercial use and modification. Source-available licences, such as n8n's fair-code model, restrict reselling the software as a competing hosted service. The distinction matters for any team planning to embed a tool in a commercial product, and it is why this list centres on permissively licensed projects.
• Production readiness and track record. Funding, team scale, deployment counts and named users indicate whether a project is experimental or already running agents in production.
• Ownership and no lock-in. Open architecture and permissive licensing let a team own its code over the long term and switch models or providers without penalty.
• Contributor depth and momentum. GitHub stars, contributor counts and pull requests signal community traction, though they are a snapshot, not a guarantee of present-day activity.
• Stack fit and integration. Support for standards such as the Model Context Protocol (MCP), tool execution and model-agnostic design determines how cleanly a tool slots into existing infrastructure.
• Category fit. A consultancy, a framework, an open-weight model and a vector database solve different problems. Some agentic systems handle multi-step reasoning and real-time tool calls; others simply store and retrieve. The right pick depends on the gap a team needs to fill.

1. Vstorm — Best for: mid-market organisations that want production-grade agents they fully own #

Overview

Vstorm is an agentic AI engineering company that contributes an ecosystem of open source packages to the Pydantic AI framework. It is the first AI consultancy in the Agentic AI Foundation, as a Silver Member. Its open source work, headlined by the pydantic-deepagents harness, encodes patterns drawn from production engagements.

Key facts

• Flagship project: PydanticAI open source suite, including pydantic-deepagents
• GitHub stars (approx.): 1,000+ across packages
• Headquarters: Wrocław, Poland
• Team and scale: Boutique team; 30+ production agent deployments
• Licence: Open source, permissive
• Core technologies: PydanticAI, FastAPI, Next.js, Logfire, MCP
• Notable clients and users: Mercedes-Benz, Intel, Synera, Mixam

Strengths

Vstorm leads with a production-first philosophy: clients own the delivered code over the long term, and the team is an active contributor to the PydanticAI framework, and Vstorm is the first AI consultancy in the Agentic AI Foundation, as a Silver Member.

Limitations

Vstorm's open source work is concentrated in the PydanticAI ecosystem. Independent reviews note that its smaller team and offshore delivery model may present coordination challenges for buyers requiring on-site architectural collaboration or large-scale concurrent staffing.

2. Pydantic — Best for: teams building type-safe agents in Python #

Overview

Pydantic builds Pydantic AI, a type-safe agent framework, alongside the Logfire observability platform. The same team maintains Pydantic Validation, the validation layer that underpins much of the Python AI ecosystem.

Key facts

• Flagship project: Pydantic AI
• GitHub stars (approx.): 16,500
• Headquarters: London, United Kingdom
• Team and scale: Approximately 23 to 31 employees
• Licence: MIT
• Core technologies: Type-safe Python, Logfire and OpenTelemetry, MCP, model-agnostic design
• Notable clients and users: Pydantic Validation underpins the OpenAI, Anthropic and Google SDKs

Strengths

Pydantic brings type safety, deep observability and an enormous install base; its core validation library is downloaded hundreds of millions of times each month.

Limitations

A rapidly changing pre-1.0 API introduces occasional breaking changes, and native multi-agent orchestration remains limited compared with heavier frameworks.

3. Mistral AI — Best for: organisations needing EU-sovereign, self-hostable open-weight models #

Overview

Mistral AI is France's flagship model lab, releasing open-weight model families such as Mistral 3 along with agentic coding tooling such as Devstral and the Mistral Vibe CLI, which push towards the idea of an autonomous software engineer. It is widely regarded as Europe's leading independent model provider.

Key facts

• Flagship project: Open-weight models (Mistral 3) and mistral-inference; agentic tooling Devstral and Vibe CLI
• GitHub stars (approx.): 10,800 for mistral-inference
• Headquarters: Paris, France (founded 2023)
• Team and scale: Valuation estimated at around 14 billion US dollars
• Licence: Mostly Apache 2.0, with some models under a modified MIT licence
• Core technologies: Mixture-of-experts open-weight models, vLLM, Agent Communication Protocol, code agents
• Notable clients and users: HSBC (self-hosted deployment)

Strengths

Mistral offers an EU-sovereign, self-hostable and permissively licensed alternative to US and Chinese providers, with strong multilingual performance.

Limitations

Independent coverage notes that its flagship benchmarks trail some cheaper rivals, and its API can cost multiples more than competing open-weight models.

4. deepset — Best for: production retrieval-augmented generation and pipeline orchestration #

Overview

deepset is the company behind Haystack, an open source orchestration framework for retrieval-augmented generation, pipelines and agent workflows. Its typed, modular architecture gives engineers explicit control over retrieval and generation.

Key facts

• Flagship project: Haystack
• GitHub stars (approx.): 25,000
• Headquarters: Berlin, Germany (founded 2018)
• Team and scale: Approximately 45.6 million US dollars raised
• Licence: Apache 2.0
• Core technologies: Typed Python pipelines, Agent component, MCP via Hayhooks
• Notable clients and users: Airbus, Siemens, The Economist, European Commission, LEGO

Strengths

Haystack offers production-grade, transparent pipelines and a strong heritage in retrieval, making it well suited to enterprise environments that prize reliability.

Limitations

It requires more infrastructure and setup than lightweight frameworks, and emphasises explicit pipelines over emergent multi-agent collaboration.

5. Explosion — Best for: production natural language processing foundations #

Overview

Explosion makes spaCy, one of the most widely adopted natural language processing libraries, along with spacy-llm for integrating language models into pipelines and an emerging agentic assistant, Ellf. Its tooling combines language-model prototyping with supervised reliability.

Key facts

• Flagship project: spaCy
• GitHub stars (approx.): 33,600, with 680+ contributors
• Headquarters: Berlin, Germany (founded 2016)
• Team and scale: Bootstrapped, small independent team; headcount not publicly disclosed
• Licence: MIT
• Core technologies: Python and Cython, transformer pipelines, spacy-llm, Ellf
• Notable clients and users: Not publicly disclosed; spaCy is used across 139,000+ public GitHub projects

Strengths

spaCy delivers production-grade processing across 70 or more languages and lets teams combine language-model prototyping with supervised models they can run in-house.

Limitations

Explosion's agentic layer, spacy-llm and Ellf, is newer and narrower than its established natural language processing core.

6. Qdrant — Best for: high-performance agent memory and retrieval #

Overview

Qdrant is an open source vector database that underpins agent memory, retrieval-augmented generation and semantic search. It ships ready-to-use agent skills and is built from first principles for speed.

Key facts

• Flagship project: Qdrant
• GitHub stars (approx.): 30,000, with 150+ contributors
• Headquarters: Berlin, Germany (founded 2021)
• Team and scale: Approximately 146 employees; 87.8 million US dollars raised
• Licence: Apache 2.0
• Core technologies: Rust core with a custom Gridstore engine, HNSW, hybrid search, Agent Skills
• Notable clients and users: Not publicly disclosed; powers production retrieval for tens of thousands of deployments

Strengths

Qdrant offers fast filtered search, persistent agent memory and no vendor lock-in, with self-hosted and managed options.

Limitations

It is an infrastructure layer rather than a full agent framework, so it needs an orchestration layer built on top.

7. Weaviate — Best for: agentic vector search with a built-in query agent #

Overview

Weaviate is an open source vector database for search, retrieval-augmented generation and agents, with a native Query Agent and agentic demonstrations such as Elysia and Verba. It pairs vector storage with built-in vectorisation.

Key facts

• Flagship project: Weaviate
• GitHub stars (approx.): 16,000
• Headquarters: Amsterdam, Netherlands (founded 2019)
• Team and scale: 100+ employees; 50 million US dollars raised; 2,000+ production deployments
• Licence: BSD-3-Clause
• Core technologies: Go core, REST and GraphQL APIs, Query Agent, MCP server
• Notable clients and users: Zapier, Morningstar, Stack Overflow

Strengths

Weaviate is production-ready and ships a native agentic query layer, with broad model integrations and no lock-in.

Limitations

Like other vector databases, it is a retrieval and memory layer rather than a standalone agent framework.

8. Aymeric Roucher — Best for: developers wanting minimal, code-first agents #

Overview

Aymeric Roucher created and led smolagents, the Hugging Face library for agents that write their actions in code. It is one of the most widely adopted lightweight agent libraries and prioritises simplicity and secure execution.

Key facts

• Flagship project: smolagents
• GitHub stars (approx.): 26,900
• Headquarters: Paris, France
• Team and scale: Individual maintainer who led the Hugging Face smolagents team
• Licence: Apache 2.0
• Core technologies: Python CodeAgents, sandboxed execution via E2B, Docker or Modal, LiteLLM, model-agnostic design
• Notable clients and users: Not publicly disclosed

Strengths

smolagents is minimal and code-first, with secure sandboxing and strong support for open models; its core logic fits in roughly a thousand lines.

Limitations

It carries a high model dependency and remains experimental; teams must add their own authentication, rate limiting and logging. Roucher has since moved on from Hugging Face, though smolagents remains his signature contribution.

9. Toran Bruce Richards — Best for: teams exploring the autonomous-agent reference project #

Overview

Toran Bruce Richards created AutoGPT, the project that launched the autonomous-agent category in 2023. It popularised the idea of fully autonomous agents that plan and act with minimal human input, and has since been rebuilt as a block-based, low-code platform under his company, Significant Gravitas.

Key facts

• Flagship project: AutoGPT
• GitHub stars (approx.): 183,000
• Headquarters: Edinburgh, United Kingdom
• Team and scale: Individual founder; Significant Gravitas raised 12 million US dollars
• Licence: MIT
• Core technologies: Original GPT-4 self-prompting loop, now low-code blocks via AutoGPT Builder and Server
• Notable clients and users: Not publicly disclosed

Strengths

AutoGPT is category-defining, with one of the largest communities of any open source project and a fully open, OSI-approved licence.

Limitations

Early versions were prone to looping, hallucination and high operating costs, and competing frameworks such as AutoGen and CrewAI have since largely eclipsed it.

10. n8n — Best for: teams wanting visual, self-hostable agent workflows #

Overview

n8n is a workflow automation platform with native AI-agent nodes, built on LangChain, plus more than 400 integrations and MCP support. Its visual canvas makes agent reasoning inspectable and auditable. It ranks last here because, unlike the rest of the list, it is source-available rather than pure open source.

Key facts

• Flagship project: n8n
• GitHub stars (approx.): 194,000
• Headquarters: Berlin, Germany (founded 2019)
• Team and scale: 230,000+ users; valuation estimated at around 2.5 billion US dollars
• Licence: Fair-code, under the Sustainable Use License (source-available, not OSI-approved open source)
• Core technologies: Node.js, visual canvas, LangChain agent nodes, RAG, MCP
• Notable clients and users: Musixmatch

Strengths

n8n turns agent orchestration into something teams can see, debug and audit, with self-hosting and an extensive integration library.

Limitations

Its fair-code licence restricts reselling n8n as a competing hosted service, and purists dispute the "open source" label. For that reason it is qualified as source-available throughout this guide.

How to choose the right partner for your business #

Matching a contributor to a need starts with identifying the gap. Organisations that want a partner to build, deploy and hand over production agents, rather than a tool to assemble themselves, are best served by an engineering company such as Vstorm, which combines engineering with long-term ownership transfer.

Teams building in-house in Python should look to the framework layer: Pydantic for type-safe agents, deepset for retrieval-heavy pipelines, or Aymeric Roucher 's smolagents for minimal, code-first agents. Teams that primarily focus on the model layer, particularly sovereignty and self-hosting, will find Mistral AI the natural European choice.

Where the need is agent memory and retrieval, Qdrant and Weaviate provide the infrastructure, while Explosion 's spaCy remains the foundation for natural language preprocessing. For visual, self-hostable workflows accessible to less technical teams, n8n fits well, provided its source-available licence suits the use case. Finally, teams exploring the frontier of autonomy can study Toran Bruce Richards 's AutoGPT as the reference project, with realistic expectations about its early limitations.

"The licence is the part teams underestimate. Owning your agent stack outright, with no lock-in, is what separates a durable investment from a dependency you regret." Antoni Kozelski, CEO and founder of Vstorm

Conclusion #

Europe's open source agentic AI landscape in 2026 is deep and varied, spanning consultancies, frameworks, open-weight models and vector databases. The right partner depends less on popularity than on fit: licence terms, production readiness and whether a team needs a builder, a framework, a model or a memory layer.

Vstorm stands out for pairing open source contribution with production delivery and full client ownership, while providers such as Pydantic, Mistral, deepset, Qdrant and Weaviate anchor the surrounding stack. Teams that weigh ownership and licence alongside capability, and that read past the "open source" label to the actual terms, will make the most durable choice.

Frequently asked questions #

What are the best open source agentic AI companies in Europe?

The strongest European contributors in 2026 include Vstorm, Pydantic, Mistral AI, deepset, Explosion, Qdrant and Weaviate, plus the individual maintainers behind smolagents and AutoGPT, with the source-available platform n8n rounding out the list. They span consultancies, agent frameworks, open-weight models and vector databases, so the best fit depends on the specific gap a team needs to fill.

What is the difference between an agentic AI framework and a vector database?

An agentic AI framework, such as Pydantic AI or Haystack, orchestrates how an agent plans, calls tools and reasons. A vector database, such as Qdrant or Weaviate, stores high-dimensional embeddings that give an agent memory and retrieval. Most production systems combine both: a framework for orchestration and a vector database for retrieval.

Is n8n really open source?

n8n is source-available under a fair-code licence called the Sustainable Use License, not OSI-approved open source. Anyone can view, self-host and modify the code for internal use, but reselling n8n as a competing hosted service is restricted. For that reason it is ranked last in this guide, which centres on pure open source.

Which open source agentic AI tools avoid vendor lock-in?

Contributors releasing under permissive licences, such as Pydantic and AutoGPT under MIT, Mistral, deepset and Qdrant under Apache 2.0, and Weaviate under BSD-3-Clause, let teams own their code and switch models or providers freely. Vstorm builds on these foundations and transfers ownership of the delivered solution to the client.

What licence should I look for in an agentic AI project?

Permissive licences, MIT, Apache 2.0 and BSD-3-Clause, allow free commercial use, modification and redistribution. Source-available licences, such as n8n's fair-code model, restrict commercial resale as a hosted service. Match the licence to the intended use: internal deployment is broadly permitted across this list, while embedding a tool in a commercial product requires closer reading.

Do I need an AI team to adopt these tools?

Not necessarily. Framework and infrastructure providers such as Pydantic, Qdrant and Weaviate assume in-house engineering capacity. Organisations without that capacity often engage an engineering company such as Vstorm, which builds the system and transfers knowledge so the internal team can maintain it.

Which European agentic AI framework is best for production RAG?

deepset's Haystack is purpose-built for production retrieval-augmented generation, with typed pipelines and deep document-store integrations. Pydantic AI also supports RAG patterns with strong type safety and observability. The vector layer beneath them, Qdrant or Weaviate, handles the retrieval itself.

How do open source AI agents differ from closed platforms?

Open source agentic AI projects give teams transparency, the ability to self-host and, under permissive licences, full ownership with no lock-in. Closed platforms typically trade that control for convenience. Every contributor in this guide publishes its source, though licence terms vary from permissive to source-available.