Grok Build Goes Open Source After Privacy Backlash, But the Repo Won't Accept Your Code

Grok Build has been at the center of a storm over the past couple of days.

Let's first recap the whole incident.

A traffic analysis targeting grok 0.2.93 showed that Grok Build uploaded the entire Git repository during testing, including files not read by the agent and the Git history. The researcher also found a test .env string in the uploaded content.

This analysis indeed proved data transmission, server-side reception, and storage, but did not prove that xAI necessarily used this data to train models.

It is normal for cloud-based AI coding tools to send the code required for a task to the model.

The point of contention is that the scope of code transmitted far exceeded the needs of the current task, and the installation and quick-start documentation at the time did not clearly explain the full-repository upload mechanism.

That is what upset everyone and drew widespread condemnation.

SpaceXAI subsequently announced the /privacy command in the Grok Build CLI as a response. This command is used to view or toggle privacy and data retention status.

Image source: Screenshot of SpaceXAI's official response

When you type /privacy in the command line, the interface clearly indicates that code will not be used for model training.

Image source: Screenshot of Grok Build CLI

A word of caution here: Not being used for training does not mean data never leaves your machine.

A cloud model needs to receive some context to understand the code. The earlier traffic analysis also found that in the version tested at the time, full-repository uploads still occurred even after turning off "Improve the model."

Then, today, July 15th, SpaceXAI announced that Grok Build would be open-sourced, along with a reset of all user quotas.

Image source: Screenshot from SpaceXAI's official account

Elon Musk also responded directly to the matter.

Image source: Screenshot of Elon Musk's public response

This open-sourcing looks like an invitation for open-source enthusiasts to participate. From another angle, it also clearly carries the color of crisis response and trust repair.

The timing between the backlash and the PR fix is just too close.

Default opt-out, historical data deletion, source code disclosure, and quota resets all appeared simultaneously. I, for one, don't believe this has nothing to do with the preceding privacy controversy.

Many people, upon seeing Grok Build open-sourced, easily misunderstand it as the Grok model also being open. That is not the case.

The official repository has open-sourced Grok Build's Rust CLI/TUI, agent runtime, and tool framework, with first-party code under the Apache 2.0 license. Grok 4.5's model weights, training code, and xAI cloud services were not opened along with it.

Image source: Grok Build homepage

Strictly speaking, this is genuine open source. Apache 2.0 permits use, modification, distribution, and commercial derivatives.

And it's not just throwing out an empty shell. You can build the CLI from source and inspect the implementations for file reading, command execution, sandboxing, tool invocation, and uploading.

The official documentation also supports configuring custom models and base_url, so you can connect this harness to other APIs, or to a locally running inference service with a compatible interface.

I took a look at Grok Build's open-source repository, and the official stance is already written very clearly:

Image source: Screenshot of Grok Build GitHub repository

• The repository currently has only one historical commit;
• The official statement explicitly says external Pull Requests and unsolicited security patches are not accepted;
• There is no official Release on GitHub corresponding to the source code repository;
• Whether the open-source code corresponds verbatim to the official binary downloaded by users still requires reproducible builds, version tags, and signing mechanisms to prove.

The official CONTRIBUTING.md states this very directly.

So my understanding of Grok's current approach is this: You can look, you can compile locally, you can Fork, and you can make your own version based on Apache 2.0;

but the official team is not ready to develop the main repository together with the community for now.

This also explains why the repository hasn't opened the usual public Issue and PR processes.

Security issues have a separate SECURITY.md channel, but ordinary feature suggestions and code contributions cannot enter the main repository for the time being.

To put it bluntly: The code is there for you to see, Fork all you want, but the official team doesn't want to accept your shit code for now...

Grok Build's open-sourcing is not a first; Codex CLI and Gemini CLI came before it. They similarly did not open the underlying model weights; what was opened were the Agent CLI and runtime frameworks.

Codex CLI:

Image source: Screenshot of OpenAI Codex GitHub repository

Gemini CLI:

Image source: Screenshot of Google Gemini CLI GitHub repository

The difference mainly lies in community governance.

Both Codex and Gemini CLI retain a complete public commit history and have open Issues, Pull Requests, and discussion areas.

Gemini CLI's README even explicitly invites the community to submit code improvements.

In other words, you file Issues when you need to, and submit PRs when you need to.

As for Anthropic's Claude Code, the source code was indeed dug out by people, and I believe basically everyone has a copy by now.

So, Grok Build has indeed been open-sourced this time, but as it stands, they don't seem to want to build a community.

However, I saw that on a certain well-known social platform, almost everyone was singing praises, with very few mentioning the fact that the official team does not accept public PRs. This is one of the few comments I saw that directly pointed out the issue.

Image source: Screenshot of a public comment on a social platform

If you are using Cursor, these past few days have indeed been a sweet moment.

Cursor directly doubled the model usage quota. Grok Build's side subsequently chose to directly reset user quotas.

Elon Musk's series of moves indeed sounds very tempting, open-sourcing the code on one hand and resetting quotas on the other, and Cursor directly doubled the quota on its side.

So I am even more firmly convinced that this open-sourcing is indeed a PR victory. It quickly diverted public attention.

Therefore, the greatest value of this open-sourcing, I think, is actually shifting a portion of trust from a corporate promise into inspectable source code.

But it can neither erase the previously captured upload behavior, nor can it independently prove that the current official binary has completely stopped full-repository uploads.