"Open source" gets thrown around loosely in AI, and Anthropic's decision to publish Claude Code's source on GitHub is the latest example of why the label matters.
The repository at anthropics/claude-code now shows the full codebase behind Anthropic's terminal-based coding agent - the tool that reads your project, executes tasks via natural language, and handles git workflows. With 108k GitHub stars and over 17,000 forks, it's clearly one of the most-watched AI developer tools around. But before you start planning your fork: the license file reads "All rights reserved" and points to Anthropic's Commercial Terms of Service. No MIT. No Apache 2.0. No GPL.
That makes this "source-available," not open source by any standard definition. You can read the code, study how it works, and file issues. You cannot freely modify, redistribute, or build competing products from it without Anthropic's permission.
What You Can Actually See
The codebase breaks down to roughly 47% Shell, 29% Python, and 18% TypeScript. That mix reveals something about the architecture: Claude Code is heavily built around shell orchestration and scripting rather than being a pure TypeScript application. The repo includes the plugin system, IDE integrations, and the GitHub @claude mention feature.
For developers curious about how agentic coding tools actually work under the hood - how they parse codebases, decide which files to read, manage context windows (the amount of text the AI can process at once), and chain tool calls together - this is genuinely useful reading material. There are very few production-grade AI agents where you can inspect the actual implementation.
The Source-Available Trend
Anthropic is following a pattern set by companies like MongoDB, Elastic, and HashiCorp: publish the code for transparency and community goodwill, but keep legal control over commercial use. It's a pragmatic middle ground. Developers get to audit the tool they're running in their terminals. Anthropic keeps competitors from repackaging their work.
The practical impact for most Claude Code users is minimal - you were already running this tool without seeing its internals. But for developers evaluating whether to build their workflow around Claude Code, being able to read the source adds a layer of trust. You can see exactly what data it collects (usage feedback and conversation data, with limited retention, and notably not used for model training according to the repo). You can verify security claims. You can understand why it behaves the way it does when it makes unexpected decisions about your codebase.
Just don't call it open source.