Is Claude Open Source? Closed Model, Open Ecosystem

No — Claude is not open source. Anthropic's models are proprietary, their weights are closed, and you cannot download or self-host them. The reason people keep asking "is claude open source" is that a lot of the ecosystem built around Claude genuinely is open, under real open-source licences.

Licence status of each component verified 31 July 2026 against the repositories themselves and Anthropic's model documentation.

That split is the whole answer to "is claude open source". The brain is closed. Much of the plumbing is open. Once you separate those two things, every follow-up question gets easy.

This piece is deliberately precise, because licensing is one of those areas where "roughly right" is wrong. Every licence claim below was checked against the actual repository rather than repeated from a summary.

Is Claude open source? The models are not

Anthropic has never published the weights of any Claude model. There is no download, no Hugging Face repository, no community edition. You reach the models only through Anthropic's apps, its API, or cloud resellers.

Claude is a family of large language models made by Anthropic, a commercial AI lab. If you want the background on the company and its funding, who owns Claude AI covers it, and what is Claude AI covers the product itself.

The models — Opus, Sonnet, Haiku — are proprietary. Anthropic has never published the weights of any Claude model. There is no download link, no Hugging Face repository, no torrent, no "community edition". You reach Claude through the Claude apps, through Anthropic's API, or through cloud resellers such as Amazon Bedrock and Google Vertex AI. In every case the model runs on someone else's hardware and your prompt travels to it.

So when someone asks "is claude open source", the honest answer about the model is a flat no, and it is not a close call.

What "open weights" actually means

Open weights means you can download the model file and run it yourself, as with Llama, Mistral, or Qwen. That still is not open source, since most such releases restrict usage and withhold training data. Claude offers neither.

The phrase people usually reach for is open weights, and it is worth pinning down because it is not the same as open source.

A model's weights are the billions of learned numbers that make it work. A model has open weights when you can download that file and run it on your own machine or your own servers. Meta's Llama family, Mistral's models, DeepSeek, Qwen and OpenAI's gpt-oss releases are all open-weights models. You can pull them down, run them offline, fine-tune them, and inspect the numbers.

Note the subtlety: open weights is not automatically open source either. Most open-weights releases ship the weights under a custom licence with usage restrictions, and they almost never ship the training data or the full training code. True open source in the traditional sense — everything needed to rebuild the thing, under an OSI-approved licence — is rare in frontier AI.

Claude sits outside both categories. No open weights, no open source, and no self-hosting.

Why Anthropic keeps the weights closed

Anthropic gives two reasons: published weights cannot be recalled and their safety training can be fine-tuned away, and API revenue funds the training runs. Report those as the company's stated positions, not settled fact.

Anthropic's stated position rests on two pillars, and it is worth reporting them as their position rather than as settled fact.

The safety argument: once weights are public they cannot be recalled, and any guardrails trained into the model can be fine-tuned back out by anyone with a GPU and a weekend. Anthropic treats model weights as the asset most worth protecting, and its published security work describes protecting them against well-resourced attackers.

The commercial argument is the obvious one. Anthropic is a company. Training a frontier model costs an enormous amount, and API and subscription revenue is how that gets paid for. Giving away the weights removes the business.

You do not have to agree with either argument. But those are the reasons, and neither shows any sign of changing.

What around Claude actually is open source

The Model Context Protocol, the official Python and TypeScript SDKs, and the Agent Skills format are all genuinely open. The models, the apps, and the Claude Code CLI itself are not.

Here is where the confusion comes from — and where the answer gets genuinely interesting. Anthropic ships a lot of open code. Just not the model.

The Model Context Protocol. MCP is the open standard Anthropic created for connecting AI models to tools, data and services, then handed to the wider ecosystem rather than keeping proprietary. The spec is public, anyone can implement it, and it has been adopted well beyond Anthropic. On licensing, be precise: the MCP project is mid-transition from the MIT license to Apache 2.0. New code and spec contributions are Apache 2.0, documentation is CC-BY-4.0, and older contributions whose authors have not consented to relicensing remain under the MIT license. The Python SDK is still plain MIT at the repo level. Either way, both are real open-source licences, and you can build on them without asking anyone. Anthropic's own announcement of the protocol lays out the intent. If you want the practical side, Claude Code MCP walks through wiring servers up.

The official SDKs. anthropic-sdk-python and anthropic-sdk-typescript are both MIT licensed — checked directly against the LICENSE file in each repository. MIT is about as permissive as it gets: use it commercially, fork it, ship it, just keep the copyright notice. Worth being clear about what that buys you, though. The SDK is an open-source client for a closed service. Free code, metered model. You can read every line of the SDK, fork it, and ship your fork — but each call it makes still bills against your Anthropic account.

Agent Skills. The Agent Skills format — folders of Markdown instructions plus scripts that Claude loads on demand — is published as a spec in Anthropic's public skills repository, alongside example skills. The repo is a mix, and the README says so plainly: many of the example skills are Apache 2.0, while the document skills that power Claude's file-creation features (docx, pdf, pptx, xlsx) are described by Anthropic as source-available, not open source. That is an honest distinction and a useful one to copy.

Claude Code. This is the one people get wrong most often, so read the licence rather than the vibes. The anthropics/claude-code repository on GitHub is not the CLI's source code. It holds the README, changelog, examples, plugins, devcontainer config and the public issue tracker. The actual CLI ships as a compiled npm package. And the repo's own LICENSE.md reads, in full: "© Anthropic PBC. All rights reserved. Use is subject to Anthropic's Commercial Terms of Service." You can read that licence file yourself in about 10 seconds. So Claude Code is not open source — it is a proprietary tool with a public repository around it. We go deeper on this in Claude Code source code, and Aider vs Claude Code sets it against a terminal agent that genuinely is open source under Apache-2.0, which makes the licence gap concrete.

Three words people constantly mix up

Open source means you may use, modify, and redistribute the code. Source available means you may only read it. Open standard means a published spec anyone can implement. Claude Code is none of the three.

Getting these straight resolves most of the argument:

1. Open source — the source is published under a licence that lets you use, modify and redistribute it. The MIT license and Apache 2.0 both qualify. The Anthropic SDKs qualify.
2. Source available — you can read the code, but the licence restricts what you may do with it. Anthropic's document skills are explicitly in this bucket.
3. Open standard — a published specification anyone may implement, independent of any one implementation's licence. MCP is the example.

Claude Code is none of the three. It is a closed binary with a public bug tracker.

The component-by-component table

The split runs clean along one line: everything that touches the model itself is proprietary, and everything that talks to the model over a wire is open under MIT or Apache 2.0.

What "is claude open source" means for you in practice

Four consequences follow. You cannot self-host Claude, you cannot run it offline, your data leaves your network on every request, and you can still build freely on the open protocol and SDKs.

Strip away the licence lawyering and the answer to "is claude open source" has four practical consequences.

You cannot self-host Claude. Not on your laptop, not on your own GPU cluster, not in your own datacentre. There is no enterprise on-prem tier that gives you weights. If a vendor claims to sell you a self-hosted Claude, that is not a real product.

You cannot run Claude offline. No network, no Claude. Everything routes to Anthropic's servers or to a cloud partner hosting the model.

Your data leaves your network. For most teams that is fine — Anthropic publishes its data-handling terms, and enterprise agreements cover retention and training. But for air-gapped environments, classified work, or regulators who forbid sending data to a third-party inference provider, the closed model is a hard architectural blocker, not a procurement negotiation.

You can build on the open parts freely. The protocol and SDKs are genuinely yours to use. Nothing stops you writing an MCP server that Claude, or any other MCP-speaking model, can call.

In our own workflow that split is a daily reality rather than a theory. This blog is built and maintained in Claude Code on native Windows, and the code that talks to Claude is our own TypeScript sitting on top of the MIT-licensed SDK — we can read and fork every line of that client, yet each generation still meters against our Anthropic account. Feeling the "free code, metered model" boundary firsthand is what made the whole licensing picture click for us: the question "is claude open source" stops being fuzzy the moment you are shipping against the open interfaces but paying for the closed brain behind them.

If you truly need local inference, the answer is not Claude — it is an open-weights model from that category (Llama, Mistral, Qwen, DeepSeek, gpt-oss and friends), run through something like Ollama or vLLM. Interestingly, the open plumbing means you can sometimes keep the tooling while swapping the brain: Claude Code with Ollama covers routing a local model behind the familiar interface, with the honest caveats about quality.

Can I see how Claude works?

Partly. Anthropic publishes research, interpretability work, model cards, and a public system prompt, so you can understand the behaviour in detail. What you never get is the artifact itself.

Partially, and more than you might expect — just never at the level of weights. This is the softer version of "is claude open source", and it has a softer answer.

Anthropic publishes a lot: research on Constitutional AI, interpretability work that pokes at what individual features inside the network represent, model cards with evaluation results and safety testing, and a public system prompt for the Claude apps. Currently the newest models — Claude Opus 5 and Claude Sonnet 5 among them — each ship with documentation describing capabilities and limits, and Claude models explained breaks the family down.

What you get from all that is understanding. What you do not get is the artifact. Reading a model card tells you how Claude behaves; it does not let you run Claude. That gap is exactly the gap between "documented" and "open source", and it is the gap that trips up the question in the first place.

The short version

The models and the Claude Code CLI are proprietary; the Model Context Protocol, the SDKs, and the Agent Skills format are genuinely open. Closed brain, open interfaces is the shape of the whole thing.

Is Claude open source? No. The models are proprietary, the weights are closed, and self-hosting is not on the menu at any price. Claude Code is proprietary too, despite the public repo — read the licence, not the GitHub badge.

But the ecosystem is a different story, and a genuinely good one. MCP is an open standard Anthropic gave away. The SDKs are MIT. The Agent Skills format is public. That combination — closed model, open interfaces — is the shape of the thing, and it is why the question feels ambiguous when the underlying facts are not.