Grok Build Turns a Prompt Into Software That Ships

Grok Build is not a chat that writes snippets.

It is the harness that plans, edits, tests, publishes, and keeps working after you look away.

That is the product now. A coding agent with a terminal, a plan mode, sub-agents, workflows, plugins, and a path from a sentence to a live app. The model matters. The harness is what made the model usable as work instead of as a parlor trick.

Build Mode started as a closed experiment: describe a site, a game, a dashboard, and watch it appear in the conversation. It was limited. Then it left the lab.

Grok Build opened to every plan on web and mobile. You describe the thing. It builds a working version in the thread. You publish it to a grok.me address. You choose who can open it. Other people can remix it. The app can call Grok models after it ships, so a trivia game can write new questions and a support page can answer in your voice.

That is the consumer face. The developer face is harder and more important.

The developer product is a coding agent that does not start by touching files.

Plan Mode writes the approach first. You approve it, cut it, or rewrite it. Then execution begins. Sub-agents take slices of the job in separate work trees. Workflows fan a task across many agents, check the results, and report back in one background run. Headless mode wires the same agent into scripts and pipelines.

This is the difference between “write me a function” and “own this change until it compiles, tests, and can be reviewed.”

The stack around that agent filled in fast.

MCP tools can pause, ask for missing input, and resume instead of dying. Plugins reach MongoDB, Vercel, Sentry, Chrome, Cloudflare, Wix. GitHub export exists. Secrets and connectors exist. Cursor now sits in the same family, which is not a branding note. It is how the coding surface and the agent harness are being pulled onto one bench.

Grok 4.6 is the model underneath a lot of the long agent work: coding, visual builds, jobs that run longer than a polite chat turn.

The recent shipping is not a keynote. It is the unglamorous layer that decides whether people stay.

Hooks that stay silent when they succeed. Workflows you can pause. Voice dictation that inserts at the cursor. Full command output instead of a chopped tail. Tables that copy as real data. Scheduled tasks that remember to check themselves. Ghost suggestions that shut up when the next line is not obvious.

That is how a coding agent stops feeling like a demo and starts feeling like a tool you will leave running.

Trust had to be rebuilt in public.

The CLI was opened after a serious data-handling failure. The harness went onto GitHub. Usage limits were reset. Local-first paths exist so inference does not have to leave the machine. Security reports go through a bounty program. None of that makes the history pretty. It does make the current tool inspectable, which a closed agent that uploads your repo never was.

A coding agent that cannot be audited will not be trusted with a real repository. That is not ideology. That is engineering hygiene.

Grok Bot is the cousin, not the same product.

Build writes software. Bot does the rest of the job on a persistent computer: inboxes, sites, handoffs, drafts you approve before they send. Faster desktop start. Better token use. iPad. File share from the phone. More languages. Switching accounts in one click.

The two products are converging on the same idea. Stop prompting. Assign the work. Verify the result.

What Grok Build is becoming is a factory floor for software.

Not a smarter autocomplete. Not a chat that pastes a file. A system that can plan a change, split the labor, run the tools, publish the artifact, and leave a trail you can reject.

It will still fail on a vague brief. It will still burn tokens. It will still need a person on anything that ships to users. The direction is not mysterious.

Chat answers.

Grok Build ships.