Grok Build

The decision comes down to two things: where you already live and how much effort you are willing to put into access.

Who it is for. For those already in Grok who want to try vibecoding without new subscriptions: the build button is part of every plan, and the finished app lives at a link right away. It also suits engineers who need an open terminal agent cheaper than Claude Code. In the Artificial Analysis index of Sep 25, 2026 a task with Grok 4.7 costs $8.82 versus $13.04 for the leader, and another model can be plugged in with your own key.

Who it is not for. Those for whom code privacy matters more than price. In July 2026 the terminal version sent whole git repositories to the vendor, until the upload was switched off on the server side.

Starting from scratch? The chat door is exactly for this: you do not need to write code at all. But to understand what to ask the agent for and how to check the result, start with the page about vibecoding — the link is below.

A good fit — section 3 shows both options. Choosing between it and Claude Code: see section 7.

What vibecoding isStarting from scratch: how to set tasks for agents with no coding experience

One agent, two faces: in chat it builds whole apps, in the terminal it works like a regular code agent.

In chat. Describe a game, website, or dashboard in words, and Grok builds a working version right in the conversation: edits are in words too, iteration after iteration. The finished app is published to a grok.me address with one button, and you choose visibility, from «only me» to «the whole internet». In the official demo, the Hyper Racer arcade game is built in 12 minutes 4 seconds.

You can remake someone else's. Any published app can be forked with the remix button and becomes your draft. Dashboards connect live data through connectors, while an integrated secrets manager stores third-party service keys.

The code is not locked in. The whole project is exported to a GitHub repository: then you can continue in your usual editor or the terminal version of the agent. You can attach your own domain to the published app.

In the terminal. A full-screen text interface, written in Rust, with the agent's code open. The signature move is parallel subagents, each in an isolated copy of the repository, so they do not trample each other's files. The rest of the set is standard for the genre: a headless mode for scripts, MCP, your own project rules in AGENTS.md.

Your choice of model. Since Sep 21, 2026 the agent has been running on Grok 4.7. In the terminal version you can also write another model with your own key into the config, if its API is shaped like OpenAI's or Anthropic's.

There are two doors: chat for everyone and the terminal for those who live in code.

Chat. Open grok.com, switch to Build mode, and write what to build. It works in the browser, on iOS, and on Android. The apps are hidden from RU stores, so the browser is more reliable on a phone.

How to install the terminal version

Step 1. Install the agent with a single command on macOS, Linux or in WSL:

curl -fsSL https://x.ai/cli/install.sh | bash

On Windows the same thing is done by a command in PowerShell:

irm https://x.ai/cli/install.ps1 | iex

Step 2. Run grok in the project folder. On the first launch it will open the browser to sign in to a Grok account; you can try it on the free plan too. For work over the API, the key is taken at console.x.ai.

Step 3. Give your first task in words and watch the process: the agent shows its steps and edits right in the interface.

The install commands and sign-in were checked on Sep 26, 2026 against docs.x.ai/build/overview and x.ai/build

An agent measurement is always a measurement of the pair «agent + model». An agent has no score on its own.

Latest measurement. In the Artificial Analysis agent index of Sep 25, 2026 the Grok Build + Grok 4.7 pair scores 56.3 points. That is fifth place out of six pairs in our AI agent index, with only the Claude Code and Codex pairs above it. The leader, Claude Code with Opus 5.5, holds 66.

Price from the same measurement. At API prices the task cost $8.82 and 39 minutes versus $13.04 and 64.5 minutes for the leader. With Grok 4.6 the same pair scored 47 points for $3.57. With Grok 4.7 the pair gained nine points, but the task became two and a half times more expensive.

Community measurement. 27,100 GitHub stars and more than five thousand forks in two and a half months of a public repository (measured Sep 26, 2026). Stars do not measure code quality, but they show well how many engineers came to take a look.

Grok Build has no separate price: it is included in Grok plans, so you pay for the plan, not the agent.

Free plan. Since Aug 19, 2026 building in chat is available on any plan, including the free one, and the vendor also calls the terminal version available for free (Sep 26, 2026). How many requests the free plan gives, the vendor does not publish.

Paid tiers. SuperGrok costs $30 a month, SuperGrok Plus $100; these are the prices the vendor shows on the plans page (Sep 26, 2026). Heavy costs about $300 according to guides. The plan's allowance is shared for the week: the agent spends it together with chat, images and voice.

We look at four things: where your project lives, what is visible in the agent's code, what has already happened and how mature it is.

The project lives with the vendor. The chat version builds and publishes apps in the SpaceXAI cloud: the grok.me address lives on its servers. There is an exit: the whole project is exported to GitHub, so the work cannot be locked inside the service. An integrated secrets manager stores third-party API keys, and they are not written into code.

You can read the agent's code, but not fix it. The repository is open under Apache 2.0, and the agent is written in Rust. This is a storefront export: the vendor does not publish development history or accept external changes. It is enough to check «what the agent does on my machine», but not for community life.

There was a code leak. In July 2026 researchers showed that the terminal version sent the entire git repository to the vendor's cloud, with history and secrets from commits. The upload was switched off on the server side, Elon Musk promised to delete everything that had been uploaded, and on July 15 the vendor opened the agent's code.

The version is no longer early. The terminal agent was announced as a beta on May 25, 2026, on Aug 7 it reached version 1.0, and by Sep 20 it had already updated to 1.0.40. An unfamiliar agent still gets its first tasks in a test project, not in a production repository.

The most frequent search question about Grok Build. The short answer: it is a trade of result against price and barrier to entry.

A different trade-off. Claude Code holds the best pair in the Artificial Analysis index, 66 points with Opus 5.5 against 56.3 for Grok Build with Grok 4.7 (Sep 25, 2026). Grok Build answers with price: the average task costs $8.82 against $13.04 at API prices.

Different entry barriers. Claude Code lives in the terminal and expects you to read code. Grok Build has a door with no threshold: chat-based building on any plan, including the free one. For someone with no coding experience, it is the only one of the two paths open right away.

Who else is nearby. Freedom to choose the model comes from OpenCode and aider: both are open and work with any provider. Grok Build also accepts another model with your own key, but out of the box it runs on Grok 4.7.

The same score, cheaper. Codex with GPT-6 Sol scores 56.7 points at $2.99 per task (Sep 25, 2026). That is almost the same score as Grok Build, three times cheaper.

The product is four months old. A short history, but a dense one: from a beta for subscribers to version 1.0, open code and free entry.