Imagine that, as a developer, you have to work your way through complex code in a repository where everything is tangled and the documentation is outdated. According to Google, this is one of the biggest problems in software development, costing a great deal of time and money. That is why it is now introducing a new product called Code Wiki, a platform that automatically creates and maintains structured documentation for code repositories. This AI-powered tool promises to help developers understand existing code more quickly.
How Code Wiki works in practice
Code Wiki works by scanning an entire code repository and generating documentation that is automatically updated after every change to the code. It is not just static text, but an interactive system that lets you jump from explanations of concepts directly to relevant files, classes, or functions in the code. For example, if you get stuck on a complex module, you can ask the integrated chat powered by the Gemini model, which takes the repository's entire current documentation into account. In addition, the tool automatically creates architecture, class, or sequence diagrams that accurately reflect the current state of the code. All of this is available at codewiki.google, where it is now in public preview and covers hundreds of public open-source projects.
Examples and benefits for developers
According to its creators, the goal is to save time. New contributors can make their first commit on day one, while experienced developers can understand new libraries in minutes rather than days. For example, in the case of Microsoft's ASP.NET Core repository, which Code Wiki has processed, the tool explains details such as distributed caching, with links directly to the code. In the chat, you can ask whether PostgreSQL can be used instead of SQL Server for distributed caching, and you will receive an answer stating that the repository does not directly support PostgreSQL, even though Microsoft's official documentation mentions this option through the IDistributedCache interface. According to Tim Anderson's article in The Register from November 17, 2025, this illustrates how AI can be accurate but may sometimes lack the full context beyond the repository.
The future and expansion for internal use
An extension for Gemini CLI, a Google tool that will make it possible to run Code Wiki locally and securely on private repositories, is coming soon. According to the announcement, this is ideal for companies where legacy code is difficult to understand because its original authors are no longer available. You can join the waiting list through developers.google.com. Omar Shams, a research engineer at Google Research and founder of Mutable.ai, which Google acquired, mentioned on Hacker News that Code Wiki is a rebuilt version of their earlier Auto Wiki project.
Potential weaknesses and comparison with other tools
In an article for The Register, Tim Anderson points out that documenting code is not easy even for AI, because models such as Gemini may not have sufficient context for a repository's entire architecture. For example, for the Vue JavaScript framework, Code Wiki generated documentation that one developer described as either too vague or too specific. It also failed to note that the vuejs/vue repository is outdated and was discontinued two years ago—there is a separate entry for vuejs/core. Anderson also notes that Code Wiki is not a true wiki because it is not collaborative or open to editing, but merely automatically generated text with diagrams. It also includes a warning that Gemini may make mistakes, so everything needs to be verified. He compares it with a similar project, DeepWiki by Devin, which is already available for private repositories as well.
Making development easier
Developers should spend their time creating, not deciphering old code, according to authors such as Rafael Marques, a product manager at Google Cloud. Code Wiki ends the era of manual and outdated documentation by offering instant understanding. If you want to try it, go to codewiki.google. According to Tim Anderson, the question remains whether AI-generated documentation is reliable enough, because it is not as authoritative as reading the code directly, and regeneration after changes can be confusing if the content changes too much from one day to the next.
