xAI spreads out all the source code of Grok Build: compile it by yourself, run it locally and offline, and even see every line of tool calls. The AI ​​programming agent circle has welcomed another blockbuster.

xAI spreads out all the source code of Grok Build: compile it by yourself, run it locally and offline, and even see every line of tool call. The AI ​​programming agent circle has welcomed another blockbuster. xAI officially announced that it will open source Grok Build. The complete source code of this encoding agent and its terminal interface (TUI) created by SpaceXAI is now available for inspection on GitHub. This is not a symbolic code disclosure, but a full exposure of the skeleton, muscles and blood vessels of the entire system to the sun. The reason given by xAI for open source is very real: making the code public is the most direct way to create a robust and reliable execution framework. In the past, you could only trust a black box to write the code for you. Now you can read all the way along the source code and see with your own eyes how the context is assembled, how the content returned by the model is parsed, and how the tool calls are distributed. Every aspect is no longer a mystery. More importantly, open source makes this framework explorable and extensible. If you are tinkering with skills, plug-ins, hooks, MCP servers or subagents, then this source code is the most authoritative instruction manual - what it accurately tells you is not the vague description in the document, but the actual implementation of how each component is loaded and called. Another thing that really excites developers is that Grok Build now supports full local-first operation. You can compile it yourself from scratch, point it to your own local inference service, and then drive all behaviors of the entire system through a config.toml configuration file. In other words, even if you don't want to rely on any cloud, you can still make this agent run clearly on your own machine. This open source code is packed with a lot of hardcore content. The first is the core logic of the agent loop, which shows how the context is assembled, how the model response is parsed, and how the tool calls are distributed. This is the heartbeat of the entire system. Second is the implementation of the toolset, where you can see how the agent reads, edits, searches code, and executes commands. The terminal interface part covers rendering, input processing, plan review and inline difference viewer, allowing people to clearly see the ins and outs of each change on the command line. Finally, there is the expansion system, which includes skills, plugins, hooks, MCP servers, and subagent access methods. It is equivalent to handing over the construction drawings for transforming this building into your hands. Project address: via AI News (author: AI Base)