Graphometer Workbench
for Grok Build
See what Grok Build is doing, and stay in control.
A local, readable window on the Grok Build coding agent, for people who direct AI agents without living in a terminal. It puts the agent's sessions, its work as it happens, and every moment it stops to ask you a question into one plain screen.
Released 2026-08-14. No further development is planned. The page, the film below, and the code stay up as they are, and the GitHub repository stays open to read, clone, and fork.
Edited from real screen captures of the running app, narrated by the founder. Nothing on screen is mocked or generated.
The five capabilities in the next section all appear in it, on screen. Everything shown is something the agent actually reported.
Read the transcript
Introducing the Workbench for Grok Build. We love Grok. We love Linux. We just don't love living in the terminal. So we built a window onto Grok Build for vibe coders that would prefer a visual representation to staring at lines in the terminal. Here's what that looks like.
The Workbench for Grok Build keeps it pretty simple. Here's the deal. You tell the agent what you want in natural language and press send. It starts thinking. Before it changes any files, it'll stop and ask for your permission and you can approve either for this one action or for the whole session. And at the end of the turn, it'll give you a little note at the bottom and tell you what it costs.
Another feature of the bench is it allows you to review and undo the agent's changes one at a time. Your own edits are never touched, so your work stays secure. Here, the agent just made three changes. This drawer on the right will list every file it touched with the real count of the lines added and removed. You can open one and see it before and after, side by side. If you don't like this change, just click undo. It'll tell you in plain words what that will do and you confirm it and that one change is gone. The other two will stay put. Here's the part that matters most. If you've edited the file yourself, undoing the agent's work never touches what you wrote, so your work stays secure.
A context meter in the upper right-hand corner shows how full the agent's working memory is at any given time. The numbers actually add up because we checked and as they get crowded, you can click compact to clear up space and see that real number drop. Below that, the privacy panel shows each setting next to its source and the one that matters that shares your data is clearly visible and asks for your permission first.
Now watch us interrupt the agent mid-sentence on purpose. There's no blank screen, no fake, everything's fine. Instead, a plain red line tells you exactly what happened. The agent was interrupted. Then the app reloads every session from the file saved on your machine. Marks it recovered and the next question answers fine, still here. So if the agent dies, you get the truth and it comes back. Never a blank screen or a fake, everything's fine.
The Workbench keeps everything in one place. You can easily rename things, run several at once. Every session on your machine is in one list, group by folder.
So that's the Workbench for Grok Build. It doesn't change what Grok can do. It makes what it already does easy to see and easy to use. Browse for a file instead of typing out the path, start a session and find it again later, rename your work, undo any changes, and always know where your privacy settings stand. The Workbench runs on your own machine, it collects nothing itself, and it's free and open source. We made Grok Build easier to build with. That's the Workbench for Grok Build.
What it does, shown in the film
Five capabilities, each demonstrated on screen. A measured note or the payoff under each one is the limit or the design decision behind it.
It asks before it changes your files
Permission cards, written in the agent's own words. In Ask mode, the Workbench shows the agent's own options and relays no approval until you pick one. The agent, not the Workbench, then writes the file. You see exactly what it wants to do before it does it.
Review and undo, one change at a time
Every changed file, with real plus-added and minus-removed counts, expandable to the full before-and-after. Undo any single change.
A context meter that adds up
The segments genuinely add up to the window. The meter reads what the agent reports, or it shows nothing at all.
It survives the agent dying
Kill the agent mid-turn and the page says the agent process is gone, then reloads every session from the files on your machine and marks it recovered.
Every session in one sidebar
All your sessions grouped by folder, several live at once on a single agent. Rename them, organize them, and move between them without losing your place.
And more
Beyond the film, the same discipline: say only what is true, and refuse visibly rather than guess.
Answer in the agent's own words
Permission, plan-approval and question cards use the agent's own option labels, verbatim. If it offers no options, the app refuses visibly rather than inventing one.
Reach for a file, not a path
Browse for a file instead of typing its path, open new sessions, and export or import whole sessions when you need to move them.
Shows the agent's own retention settings
These are Grok Build's own retention settings, which the Workbench reads back and displays next to each source, not a statement about xAI's data practices. The one changeable setting is set through the agent's own protocol, then verified by reading it back.
It runs on your machine, and collects nothing itself.
These are properties of the Workbench, checked in the code and against the recorded checks, not promises about Grok Build or xAI.
The Workbench server binds 127.0.0.1 and no other interface: its listen address, not a claim about Grok Build or xAI.
The Workbench sends nothing out: no analytics, no phone-home, no tracking of any kind.
The Workbench gates its local interface with a fresh access token each time the server starts.
The Workbench holds its page to a locked-down CSP that permits only what it needs and nothing more.
The Workbench renders every string the agent sends as text, and never runs it as markup or code.
The Workbench is open source under the MIT license, with more than 1,400 deterministic checks that run green.
These claims are about the Workbench itself. They say nothing about what Grok Build or xAI do on their side. The Workbench is the local window you run; how the agent behaves upstream is the agent's own matter.
What you need to run it
Small on purpose. The server runs TypeScript directly: no build step, no dependencies to install.
Node.js 22+. The server runs TypeScript directly:
no build step, no dependencies.Ready when you are. It's on GitHub.
The whole thing lives in one public repository under the MIT license. Read the code and run it locally. The Workbench itself needs no account and no sign-up; Grok Build needs its own login, installed separately.