Graphometer Workbench for Grok Build

A local web interface for the Grok Build CLI — built for people who direct AI agents but do not write the code themselves.

for Grok Build · not affiliated with xAI

Grok Build already ships subagents, worktrees, skills, a real context accounting system and per-change tracking — all of it invisible outside a full-screen terminal UI. Graphometer Workbench is a window onto that machinery: it adds visibility and reach, never new agent capability. Everything on screen is something the agent actually reported; nothing is invented, estimated, or silently dropped.

What it does

• Review and undo the agent's changes, per change. Every changed file with real +N/−M counts, expandable to before/after per change, with agent-versus-human attribution — Undo this change reverts the agent's edit and leaves yours untouched.
• Answer the agent in its own words. Permission, plan-approval and question cards render inline in the conversation using the agent's own option labels, verbatim — never an invented button, never an auto-approval. If the agent offers no options, the app refuses visibly rather than guessing.
• An honest context meter. The segments genuinely sum to the window; overlapping categories are listed beside the bar, not stacked into it (the naive stacking would overstate context pressure by 68%). A live mid-turn tick was built, measured dishonest on the wire, and reverted — the meter reads what the agent reports, or nothing.
• Survives the agent dying. Kill the agent mid-turn and the page narrates the death, the automatic respawn, and each session's recovery — it never shows a blank page and never claims recovery it can't verify.
• Multi-session. Every session on the machine in one rail, grouped by folder — several live at once on one agent process, each streaming to its own pane. Rename, archive, delete, export/import.
• A privacy panel that tells the truth. Your coding-data retention and zero-data retention state, each with its source and its real lever named. The one setting that is changeable is changed through the agent's own protocol request — with the result verified by reading state back, never assumed from an acknowledgment.
• Local-only, by construction. Binds 127.0.0.1 only, per-run access token, strict Content-Security-Policy, zero telemetry, zero third-party calls. Every string the agent sends is treated as hostile input: text, never markup. See SECURITY.md.

Behind that: 1,400+ deterministic checks run green, and the project's rule is that every capability claim is proven on the wire or in a live browser before it is written down. The wire captures and evidence packs those proofs live in (the docs/evidence/* paths cited in code comments) stay in the private working tree, not this repository — ask and we'll share them.

What it looks like

Watch the 3-minute launch film — narrated by the founder, edited from real screen captures of the running app.

Requirements

• Linux (tested) or Windows via WSL2 — the process-safety machinery (process-group cleanup with /proc verification) is Linux-specific by design. macOS is unverified and not claimed.
• Node.js 22+ (the server runs TypeScript directly — no build step, no dependencies, no package.json).
• The Grok Build CLI, installed and authenticated (grok --version — tested against 1.0.4). The app drives grok agent over stdio; your existing login is used, and your token is never read by this app.

Run

git clone https://github.com/graphometer/workbench-for-grok-build.git
cd workbench-for-grok-build
./start.sh

start.sh starts the server on port 4311 and opens your browser with the one-time access URL (printed in the terminal if no browser opens). Ctrl+C stops it; ./stop.sh from another terminal also works, and shuts down cleanly — the agent process tree is verified dead, and your ~/.grok/config.toml is hash-checked unchanged on the way out.

Different port: STUDIO_PORT=4318 ./start.sh

Configuration (optional)

Copy graphometer.env.example to graphometer.env next to start.sh — per-machine settings, never tracked: workspace root for the folder picker, folders the app must never list, default new-session folder, port.

Development surfaces (a raw wire view at /bridge and a Shell inspector tab) are compiled out of the product build; start with STUDIO_DEV=1 to get them.

What this app never does

• Never sends anything anywhere except your local Grok Build agent.
• Never reads your API token (two named retention booleans are the only fields that ever leave ~/.grok/auth.json's parse, server-side).
• Never writes your agent configuration casually — config.toml is hashed before and after every run and the app reports if it changed; auth.json is written only by the agent itself, only when you flip the retention switch, and the result is digest-logged.
• Never renders agent output as markup — no innerHTML anywhere in the page, enforced by checks.

Status

v1.0. Linux tested end-to-end against Grok Build 1.0.4 (and 1.0.0-1.0.3 before it). Windows via WSL2 verified as a fresh install on a real Windows 11 machine — including the shutdown behavior that the first attempt caught failing, which is why the launch suite now reproduces that exact hang against the old code. Issues and honest platform reports welcome. macOS reports especially — it remains unclaimed until someone proves it.

License

MIT · built by Grant Williams — graphometer.ai — with an AI build-and-review crew (Grok, Kimi, Codex, Claude). Graphometer is an independent project and is not affiliated with, endorsed by, or sponsored by xAI.