Grok Build Explained: Setup, Pricing, and Is It Any Good?
There's a new AI coding agent to try every few weeks. You've probably tried Claude Code, Codex, or Cursor, and now there's Grok Build, which SpaceXAI launched in May and has updated more than 40 times since version 1.0 landed on August 7.
It doesn't help that the name covers two products. One is a coding agent that lives in your terminal and edits real repos. The other is a mode inside the Grok chat app that spins up a little game or website from a prompt (fun, but it won't fix your flaky test suite).
I wanted a simpler answer: is the terminal agent worth installing, and what does it cost to run?
So I went through everything SpaceXAI has published about it: the docs, every launch post, and the changelog. I checked each price and spec against SpaceXAI's own pages as of late September 2026, then read the independent benchmark results and the reporting on its July privacy scare.
The short version: it's one of the cheapest serious coding agents you can run today, it reads your existing Claude Code setup as-is, and its privacy history is something to read about before you point it at work code.
What is Grok Build? The 30-second answer
Grok Build is SpaceXAI's terminal coding agent: it plans changes, edits files, and runs commands inside your repo while you review the work. It launched as an early beta on May 25, 2026, and it now runs Grok 4.7 by default.
(SpaceXAI is the name now on every x.ai page. If you still call the company xAI, everyone knows who you mean.)
You install it with one command, open a project folder, type grok, and talk to it in plain English. It reads your codebase, asks questions when your request is vague, proposes a plan for bigger jobs, and shows every change as a diff you can approve.
Here's the quick version:
The GitHub README sums it up as an agent that understands your codebase, edits files, runs shell commands, searches the web, and handles long-running tasks.
You can use it three ways: interactively, headlessly in scripts, or inside another editor. That third one is how Grok Build gets out of the terminal.
How Grok Build works
A Grok Build session starts the moment you run grok inside a project folder. Before you type anything, it loads any instruction files it finds (AGENTS.md, CLAUDE.md, and a few others), so your team's conventions are already in context.
Then you ask for something. A small request, like "rename this function and update every caller," usually goes straight to work: Grok Build searches the code, edits the files, runs your tests, and reports back with diffs.
A bigger or fuzzier request gets more care. If you ask for something ambiguous, Grok Build can come back with quick multiple-choice questions (pick a framework, a schema, a design direction), and your answers flow into its plan.
For the big stuff, you start in plan mode. The agent explores the repo and writes a plan, and every edit stays blocked until you approve it. You can approve the plan, comment on single lines, or send it back for changes.
How much it asks before acting depends on the permission mode you pick:
- Ask (default): prompts for anything you haven't already allowed.
- Auto (where available): a classifier approves safe tool calls, and risky ones can still prompt you.
- Always-approve: skips the prompts, though your deny rules and hooks still apply.
Once you switch it on, it also builds a memory. The feature is off by default, so start a session with grok --experimental-memory or turn it on in your config.
After each turn, Grok Build writes short notes about your project's conventions and decisions, like which command runs your tests and why plain cargo test fails the integration tests.
Later sessions read those notes back, and /memory lets you browse them (handy for finding the file to fix when a note is wrong).
How to install and access Grok Build
Getting Grok Build running takes about as long as making coffee, assuming your coffee is instant. Here's the whole process, straight from the official docs:
1. Install it with one command
On macOS or Linux, open a terminal and run:
curl -fsSL https://x.ai/cli/install.sh | bash
On Windows, open PowerShell and run:
irm https://x.ai/cli/install.ps1 | iex
Run grok --version afterward to confirm it worked. (If you'd rather compile it yourself, the source is on GitHub and builds with Rust, though Windows source builds are officially "best-effort.")
2. Open a project and start a session
Move into the folder you want to work on and start the agent:
cd your-project
grok
3. Sign in
The first time you launch it, Grok Build opens your browser so you can sign in. On a remote server or container with no browser, run grok login --device-auth instead: it prints a short code you can enter from any device.
For scripts and CI, skip the login entirely and set an API key with export XAI_API_KEY="xai-...". Usage on that key is billed at SpaceXAI's API rates.
4. Give it a first task
Two good starter prompts from the docs are "Explain this repo" and "@src/main.rs Walk me through this file." They're harmless, and they show you quickly how well it reads your code.
Who can use Grok Build
At launch in May, Grok Build was limited to SuperGrok and X Premium Plus subscribers. That's changed: the product page now says it's available to try for free.
Today there are four ways in:
- Free tier: sign in and go, with usage limits SpaceXAI doesn't publish.
- SuperGrok plans: higher limits on a paid Grok subscription.
- API key: pay per token, which suits automation and CI.
- Enterprise: team logins through your company's identity provider, admin-pinned settings, and zero data retention at the team level.
Grok Build pricing
Grok Build pricing comes down to how you sign in, and SpaceXAI doesn't publish exact usage caps for any plan. Here's what each route costs as of September 30, 2026:
Plan prices come from grok.com/plans, and the API rates from SpaceXAI's pricing docs. A few details worth knowing before you pick:
- Long prompts cost double on the API. Once a request's prompt reaches 200,000 tokens, Grok 4.7 bills $4 in and $12 out per million for the whole request. Cached input drops to $0.50 per million on shorter prompts.
- Grok 4.7 Fast costs double. It's the same model on faster hardware at twice the token rate (1.5x for long prompts), it's only offered in Grok Build and Cursor, it's billed through your plan there, and it isn't part of the free tier.
- Paid plans share one weekly usage pool across Chat, Imagine, Voice, and Build, and it resets every week. Engadget's walkthrough also found free users burn through their allowance fast.
- Grok Build 0.1 is a separate, cheaper model. It's SpaceXAI's fast coding model from May (requests to the retired grok-code-fast-1 now route to it), priced at $1 in and $2 out per million tokens for prompts under 200,000 tokens, with a 256,000-token context window. It isn't Grok Build's default anymore.
Why a cheap model can still run up a bill
Grok 4.7's token prices look low next to its rivals, but it's a chatty model. In Artificial Analysis testing, it used about 81,000 output tokens per task on Artificial Analysis's Intelligence Index at its highest effort setting, more than double what Grok 4.6 needed.
So if you're paying per token, check /usage after a few real tasks before you run Grok Build in a CI job that fires on every pull request. Dropping the effort level with /effort is the fastest way to trim the bill on routine work.
Grok Build features worth knowing
Grok Build has picked up a lot since May. The product page lists 18 features, and the changelog adds new ones almost weekly, so here are the seven that change how you'd work with it.
1. Plan mode with a real approval gate
Plan mode is the feature I'd turn on first. The agent drafts a plan file, and SpaceXAI's docs say other edits stay blocked until you approve it, even if you've set the session to auto or always-approve.
One caveat from the same docs: plan mode gates the edit tools, and shell commands still follow your permission mode. A bash redirect can still write a file, and subagents aren't edit-gated by the parent's plan mode, so don't treat it as a sandbox.
2. Subagents, worktrees, and workflows
For bigger jobs, Grok Build hands pieces of the work to subagents, child sessions with their own context that report back when they finish. The built-in types cover general work, read-only exploring, and planning, and you can define your own.
Subagents can run in their own Git worktrees, so parallel changes don't trample each other.
Workflows go further. You describe a large job in plain English, and Grok Build fans it out across up to 128 agents (1,024 for big runs), checks the results, and hands you one report.
A saved workflow becomes its own slash command, so a PR review you liked turns into /pr-review 5137 next time. /deep-research ships built in.
3. Headless mode and ACP
Run grok -p "your prompt" and Grok Build works without the interactive screen, printing plain text, JSON, or streaming JSON. That's the mode for scripts, bots, and CI pipelines.
The Agent Client Protocol (ACP) is how Grok Build shows up inside other tools. Run grok agent stdio and any editor or app that speaks ACP can drive it. That's your route into an editor for now, since SpaceXAI doesn't ship an official VS Code extension.
4. AGENTS.md, skills, plugins, hooks, and MCP servers
Grok Build reads the usual project instruction files (AGENTS.md, CLAUDE.md, and friends) plus rules folders from .grok/, .claude/, and .cursor/. Skills are folders with a SKILL.md file that turn into slash commands, and /skillify turns a session you liked into a new skill.
Plugins bundle skills, hooks, and MCP servers behind one install, and there's a built-in marketplace. Hooks run your own scripts on edits and tool calls, and MCP servers connect tools like Linear, Sentry, and Postgres.
The big one for anyone switching: SpaceXAI says Grok Build is "fully compatible with Claude Code with zero configuration needed." It picks up your existing Claude Code plugins, skills, MCP servers, hooks, and CLAUDE.md files, and grok import brings over old Claude Code sessions.
5. Memory and /goal
Memory arrived on September 16 and still has to be switched on. Once it is, Grok Build keeps markdown notes on your project's conventions and decisions, one topic per file, and reads them before touching related code. Secrets and anything your docs already cover are left out on purpose.
/goal is for handing off a longer job. Give it an objective like "Migrate the auth module to the new API," and Grok Build keeps working through a checklist until the task is done and verified. You can check in with /goal status, pause it, or clear it.
6. Agent Dashboard
Once you're running a few sessions at once, grok dashboard puts all of them on one screen. Anything waiting on you (a question, an approval) floats to the top, and you can reply without opening each session.
7. Bring your own model
Grok 4.7 is the default, and you can add any model with an OpenAI-compatible or Anthropic Messages endpoint to ~/.grok/config.toml, including a local one, and switch with /model. Composer 2.5 also sits in the model menu.
Grok Build on GitHub: open source, and why it happened
The full Grok Build source is on GitHub at xai-org/grok-build, under the Apache 2.0 license. It's written in Rust, it had about 27,000 stars as of September 30, and it includes the agent loop, the tools, the terminal UI, and the whole extension system.
Two catches: the public repo is synced periodically from SpaceXAI's internal codebase, and the README says outside contributions aren't accepted. You can read it, fork it, and build it, but you can't send a pull request upstream.
SpaceXAI announced the open-source release on July 15, pitching it as a way to see exactly how the harness works and to run it "fully local-first" against your own model.
The July privacy scare
The release came three days after a rough news cycle. On July 12, a security researcher showed that Grok Build version 0.2.93 was uploading developers' entire tracked repositories to a cloud storage bucket.
That meant full Git history and committed secrets, including files the agent never opened. The privacy toggle meant to stop data sharing didn't stop it.
The uploads stopped after a server-side change, and Elon Musk promised the previously uploaded data would be deleted. SpaceXAI also pointed users to the /privacy command, which shows your retention status and turns data retention off.
If you used Grok Build before mid-July on a repo with secrets anywhere in its history, rotate those keys.
For company code, enterprise teams can turn on zero data retention, which SpaceXAI enforces at the team level.
{{templates}}
Grok Build vs Build Mode on grok.com
Build Mode is the other product that goes by the Grok Build name.
Build Mode on grok.com is a no-code app builder inside the Grok chat app. You describe a website, game, or dashboard, Grok builds a working version in the chat, and you publish it to a grok.me link.
It launched in July as an early beta for SuperGrok Heavy. It's been available on every plan, on web and mobile, since August 19.
The two do connect: a Build Mode app can be exported to GitHub, and SpaceXAI suggests picking it up from there in the terminal agent. If you came here wanting to make a quick app without touching code, Build Mode is the one you want.
Is Grok Build any good?
Grok Build is good, great on price, and a notch below the top coding agents right now. On the Artificial Analysis Coding Agent Index, Grok Build running Grok 4.7 scored 56, nine points better than it managed with Grok 4.6.
That put it fourth among models tested in their own harnesses at launch, behind Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5. Respectable company, and a big jump for one model update, but Anthropic's and OpenAI's top setups still finish more of the hard, multi-step tasks.
What Grok Build does well
- Low cost of entry: a free tier and $2/$6 API pricing make it one of the cheapest serious agents to try.
- Near-zero switching cost: it reads your Claude Code setup as-is, so a trial doesn't mean rebuilding your config.
- Serious parallelism: subagents, worktrees, and 128-agent workflows all come built in.
- A plan review auto modes can't skip: edits wait for your approval even under auto or always-approve.
- Open source: you can read exactly how the agent works.
Where Grok Build falls short
- A trust debt: the July upload incident is recent, and some security teams will want more than a server-side fix and a promise.
- Verbose model: Grok 4.7 burns through tokens, so pay-per-token users can see bills climb on long tasks.
- Safety defaults are loose: the sandbox is off by default, and its network blocking only works on Linux.
- No first-party IDE extension: editors connect through ACP.
- Fuzzy limits: SpaceXAI doesn't publish how much each plan gets you.
Grok Build vs Claude Code, Codex, and Cursor
The short version: Claude Code and Codex both still finish more of the hard, multi-step work, and Codex is the easy pick if you already pay for ChatGPT. Cursor suits people who want an editor first. Grok Build is the cheapest open-source option of the four.
Prices come from Claude's pricing page, OpenAI's Codex pricing docs, and Cursor's pricing page. Price, benchmarks, and security are where Grok Build and Claude Code split, and all four sit in a crowded field of AI coding agents.
One thing to know before you compare the last two: Cursor is now part of SpaceX, which completed the acquisition on August 14, 2026, so Grok Build and Cursor share a parent company.
Who Grok Build is for (and who should skip it)
Grok Build is a strong fit if you:
- Want a capable terminal agent on a small budget, or free.
- Already use Claude Code and want a second agent that reads the same setup.
- Run big, parallel jobs like PR reviews, issue triage, or codebase audits, where workflows shine.
- Care about reading and modifying the agent's source, or pointing it at a local model.
Skip Grok Build (for now) if you:
- Work under strict data rules and can't get sign-off after the July incident.
- Need the strongest agent available for gnarly, multi-day refactors, in which case Claude Code is the safer bet, and a few Claude alternatives round out the field.
- Live inside VS Code and want a native extension. GitHub Copilot fits that better.
- Aren't a developer at all.
Say the work you want off your plate is triaging support tickets, chasing invoices, or filing bug reports from Slack threads. AI tools built for business work and the Microsoft Copilot alternatives are made for that kind of job, and Lindy, an AI teammate that lives in Slack, is one of them.
{{cta}}
Give Grok Build a throwaway repo first
Grok Build has earned a trial, especially at its price, and its history has earned a careful one. The safest way to find out if it fits is a low-stakes first hour:
- Point it at a repo with nothing to lose, like a side project or a fresh clone with no secrets in its history, and run /privacy to see where your data retention stands.
- Start in plan mode with a real but contained task, and see whether its plan matches what you'd have done.
- Try one headless run with grok -p on something repetitive, then check /usage to see what it cost.
If those three go well, Grok Build has earned a place in your toolkit, maybe as your main agent and maybe as the cheaper second opinion next to the one you already trust.
FAQ
What is Grok Build built on?
Grok Build is built on SpaceXAI's Grok 4.7 model by default, and the agent itself is written in Rust. You can swap in other models, including Composer 2.5 or any OpenAI-compatible or Anthropic Messages endpoint, from the config file or the /model command.
Is Grok Build free?
Yes, Grok Build is free to try with a limited free tier. Heavier use needs a SuperGrok plan ($10 to $300 a month) or an API key billed per token, and SpaceXAI doesn't publish the exact limits for any tier.
How do I access Grok Build?
You access Grok Build by installing it from your terminal with curl -fsSL https://x.ai/cli/install.sh | bash (macOS and Linux) or the PowerShell command on Windows. Then run grok in a project folder and sign in through your browser.
What is Grok Build best for?
Grok Build is best for developers who want a low-cost terminal agent for multi-file changes, parallel work like PR reviews and audits, and scripted jobs in CI. It's also a natural second agent for Claude Code users, since it reads the same config.
How much does Grok Build 0.1 cost?
Grok Build 0.1 costs $1 per million input tokens and $2 per million output tokens on SpaceXAI's API for prompts under 200,000 tokens, with cached input at $0.20. It's SpaceXAI's fast coding model from May, with a 256,000-token context window, and it's no longer Grok Build's default.
