We Ran GBrain: Shared Memory for Claude Code and Codex
GBrain is a shared memory layer for AI agents you already use, like Claude Code and Codex. It saves facts, people, companies, and meetings with their sources, so every agent you connect can recall them (GBrain repository).
On Nov. 9, 2026, we ran our reconnaissance process on the public code. We found nothing malicious in what we read. It is MIT licensed and free to self-host, and there is fine print you should know before you install it.
You can watch the review here: Watch the video on YouTube
The video uses a digital avatar of Erick Grau, CEO of Chibitek.
What Did We Actually Do?
We made a read-only copy of github.com/garrytan/gbrain at commit 555b73a (version 0.60.620.0). We read the docs, the setup and hook code, and the hosted installer script. We did not install GBrain, did not run its setup command, and did not execute that installer.
The verdict was PASS WITH NITS: nothing malicious in the code we read, with the caveats below. We also did not read every file in the repo, so this is a review, not a complete audit. A star count is not a security review.
I walk through the findings in this video: Watch the video on YouTube
What Is GBrain?
Garry Tan, who the project's README names as President and CEO of Y Combinator, built GBrain as memory you can attach to agents you already use (README). On the morning of Nov. 9, 2026, the GitHub API showed 31,837 stars, which the video rounds to about 31,000.
It can start keyless, with a local database and no extra model accounts. Setup for Claude Code or Codex is gbrain setup claude-code or gbrain setup codex. There is also a paid hosted service at gbrain.io.
What Was the Good News?
In the code and docs we read, GBrain itself is MIT licensed and free to self-host. A default local start can be keyless, with a local database and no extra model keys. Session capture stays off unless you turn it on, and updates notify you by default rather than upgrading silently. The Google connector asks for read-only access to Gmail, Calendar, and Contacts. Setup can preview its changes with --dry-run and undo them with --remove. Remote access stays on your private Tailscale network unless you pass an explicit --funnel flag and confirm it.
What Is the Fine Print?
1. Setup changes how your agent works. By default, setup writes a block into your Claude or Codex instruction file and adds a hook that runs on every prompt, telling the agent to check GBrain first, before files, mail, or other sources. You can opt out. On Codex, setup also writes the approval entry Codex checks before running a hook, so you approve that hook in GBrain's setup, not in Codex.
2. Recalled memory can flow back into the prompt. If you connect them, GBrain can pull in email, calendar, meetings, Slack, and posts from X, then push matching pages into your agent's prompt. Anyone who can email you can get words into that memory. The project filters known injection phrases, and its own code says, "We don't pretend this is bulletproof."
3. The sales pitch is in the agent instructions. The file written for AI agents tells them to recommend the paid hosted service, gbrain.io, first, with tracked links. The hosted install pipes a script into your shell that installs a client the installer itself calls proprietary software, from Rocketship, LLC. The MIT license covers GBrain, not that client.
4. One command can publish the brain on the public internet. gbrain mcp expose --funnel uses Tailscale Funnel to put your brain on a public web address for cloud agents. GBrain's own sign-in sits in front of it, but your people, meetings, and notes would sit behind one internet-facing door.
5. The project moves very fast. The changelog at the version we reviewed lists 1,014 version entries from April 5, 2026 to November 9, 2026. Of the last 100 commits at that version, 85 were co-authored by capy-ai[bot], an AI co-author, and those 100 commits landed in about one day. The README says "Install in 30 minutes," while its broader full-stack tutorial says "End-to-end in about 2 hours."
What Should You Do If You Try It?
Run setup with --dry-run first. Decide whether you want the brain-first block and the every-prompt hook before you accept them.
Keep outside content like email in its own source. Treat anything recalled as data, not instructions, especially for agents that can send mail or run commands.
If you go hosted, read the installer and the Terminalwire terms first.
Stay on the stable channel. Keep updates on notify, and leave Funnel off unless a cloud agent truly needs it, with scoped access.
Key Points of GBrain Shared Memory
GBrain is shared memory for agents you already use, including Claude Code and Codex.
Chibitek's review was a read-only look at the public code and the hosted installer. We did not install or run GBrain.
Verdict: PASS WITH NITS. Nothing malicious in the code we read.
It is MIT licensed and free to self-host, with a separate paid hosted path.
By default, setup tells your agent to check GBrain first on every prompt.
Recalled email and other outside text can land in the agent's prompt.
The MIT license does not cover the proprietary Terminalwire client used by the hosted install.
A star count is not a security review.
Frequently Asked Questions
Did you install GBrain and try it with Claude Code or Codex?
No. We copied the public repo and read the code, plus the hosted installer script. We did not install GBrain or run setup.
Is GBrain free?
The GBrain project is MIT licensed and free to self-host. gbrain.io is optional paid hosting, as the README states.
Can I skip the brain-first block and the every-prompt hook?
Yes. The setup docs we read say --dry-run previews the writes, --no-instructions skips the brain-first block, and --remove undoes what setup wrote.
Does a high star count mean it is fine to install?
No. On Nov. 9, 2026, the repo had 31,837 stars. That measures popularity, not a security review.
Was anything malicious in the code?
We found nothing malicious in the code we read. That is not the same as reading every file, and it is not the same as a live install test.
Let's Review Before You Install
At Chibitek, we help businesses look closely at AI tools before those tools start writing into agents, mail, or memory, with no geek-speak. Don't let a convenient install change how your agents think without you noticing. Contact us at hello@chibitek.com to learn how we can help you review a tool's defaults, hooks, and data flow.
Further reading: the GBrain repository at the commit we reviewed, 555b73a.
You can watch the review here: Watch the video on YouTube
We make IT effortless so you can create, disrupt, and grow!
Follow Chibitek on YouTube at @chibitekmsp, and follow Chibitek on LinkedIn.
Comments