Is Claude Open Source? Where the Line Falls

No. Claude is not open source, and neither is Claude Code. The models are proprietary and the weights aren't published.

But the question keeps getting asked because parts of Anthropic's stack genuinely are open, and one of them sits in a public GitHub repository that looks, at a glance, exactly like an open-source project. Here's the line, drawn precisely:

The models

Claude is a hosted, closed-weight model. You reach it through an API or an app; you don't download it, you can't run it on your own hardware, and you can't fine-tune the base weights yourself.

This isn't a gap waiting to be filled — it's the business model. If what you need is a model you can run locally and modify, Claude is the wrong tool no matter how the licensing evolves, and you want open-weight families instead. We covered that path in running AI locally.

Claude Code: public repository, proprietary code

This is where nearly all the confusion comes from.

github.com/anthropics/claude-code is a public repository. It has issues, releases, documentation, and thousands of stars. It looks open.

Its LICENSE.md is one line, and it reserves all rights to Anthropic PBC, with use governed by Anthropic's Commercial Terms of Service.

So: public ≠ open source. The repository exists for documentation and issue tracking. You can read it, file bugs, and watch releases. You cannot fork the product, modify it, and redistribute your version — none of the permissions an open-source license grants are granted here.

"Source-available" is the closest accurate label for the parts you can see, and even that overstates it for the application itself.

The Claude Agent SDK needs splitting apart, because the two language packages don't agree with each other:

• Python — claude-agent-sdk-python ships an ordinary MIT LICENSE. The Python source in that repository is open source: you can fork it, patch it, and vendor it.
• TypeScript — claude-agent-sdk-typescript does not. Its LICENSE.md is the same one-line all-rights-reserved notice as Claude Code, with use governed by Anthropic's Commercial Terms.

Both READMEs, and the Agent SDK docs, carry the same paragraph: use of the SDK is governed by the Commercial Terms, except where a specific component or dependency has a different license in its own LICENSE file. The Python repository's MIT file reads as that kind of exception: it covers the Python code, not everything around it.

Two things sit underneath both packages and are proprietary regardless of which you install. The first is the Claude Code binary. Both SDKs bundle it, so most installs need no separate Claude Code install, and it keeps Claude Code's all-rights-reserved license — the MIT file in the Python repository covers the Python wrapper, not the binary it drives. The second is the service that binary talks to: Anthropic's Commercial Terms govern use of Anthropic API keys, whichever SDK you chose.

So "is the Agent SDK open source?" depends on which layer you mean, and only the Python wrapper gets a yes. There are also branding restrictions on top — you may say your product is powered by Claude; you may not call it Claude Code.

What actually is open

Beyond the Python Agent SDK source, two more things, and both are more useful than they sound.

The Anthropic API SDKs are MIT-licensed. anthropic-sdk-python and its TypeScript sibling are ordinary open-source client libraries. You can read them, fork them, vendor them, patch them. They're clients for a closed service — MIT covers the wrapper, not what it talks to — but as libraries they're as open as anything else in your dependency tree.

MCP is an open standard. The Model Context Protocol is open source, and it's the piece with real strategic weight. Anthropic created it, but it's supported across a wide ecosystem — Claude, ChatGPT, VS Code, Cursor, and many more. An MCP server you write today works in any client that speaks the protocol.

That last point is the practical hedge against lock-in. Your integration work is portable even though the model isn't. If you build tooling around MCP and later switch models, the servers come with you.

Does it matter?

Depends entirely on why you asked.

• You need to self-host, audit weights, or run offline. Then yes, decisively — go open-weight.
• You need to avoid vendor lock-in. Partly. Build your integrations on MCP and keep model-specific logic thin, and switching costs stay manageable.
• You want to inspect or modify Claude Code's behaviour. You can't fork it, but the extension points are real: skills, hooks, subagents, and plugins let you change a lot without touching the source.
• You're checking whether it's free. Different question. Open source and free aren't the same thing, and Claude's plans are on Anthropic's pricing page.

For most teams the honest answer is that the licensing matters less than the exit cost, and the exit cost is mostly about how much model-specific glue you wrote — not about who owns the weights.

FAQ

Is Claude open source?

No. The Claude models are proprietary and the weights aren't published.

Is Claude Code open source?

No. Its repository is public for documentation and issues, but the license reserves all rights to Anthropic and use falls under Anthropic's Commercial Terms of Service.

Why is Claude Code on GitHub if it isn't open source?

The repository hosts documentation, releases, and issue tracking. A public repository grants no rights on its own — the license does, and this one grants none of the usual open-source permissions.

Are any Anthropic tools open source?

Yes. The Anthropic API SDKs for Python and TypeScript are MIT-licensed, the Python Claude Agent SDK's source is MIT too, and the Model Context Protocol is an open standard with broad ecosystem support.

Is the Claude Agent SDK open source?

It depends which package. The Python SDK's repository carries an MIT license, so its source is open source. The TypeScript SDK's license reserves all rights to Anthropic, with use under the Commercial Terms. Both bundle the proprietary Claude Code binary, so neither gives you an end-to-end open stack.

Can I self-host Claude?

No. Claude is available through Anthropic's API and through cloud partners, but the weights aren't distributed, so there's no self-hosted deployment.

What are the open-source alternatives to Claude?

Open-weight model families you can download and run yourself. Our guide to Claude alternatives covers both hosted and open-weight options.

The short version

Claude and Claude Code are proprietary. The API SDKs are MIT, the Agent SDK's Python source is MIT while its TypeScript package reserves all rights, and MCP is an open standard — which means the integration layer of your work is largely portable even though the model isn't. If open weights are a hard requirement, look elsewhere; if the worry is lock-in, build on MCP and keep your model-specific code thin.

If what's really driving the question is control rather than licensing, note that the thing you most want to own isn't the weights — it's the setup: what runs, where it runs, and whether you can hand it to someone else. Taku is a desktop workspace built on that idea, where AI apps and agents run on your machine and stay yours to change. Taku is in Beta, and the Mac app is available now.