Last verified

SpaceXAI · Grok Build

Grok Build: free software on a meter you cannot read

SpaceXAI's coding agent for the terminal — Apache-2.0, included on every Grok plan including the free one, and metered in a unit the vendor never publishes.

Grok Build is a coding agent you install and run in a project directory, with a full-screen terminal interface, a headless mode and an open-source harness under Apache-2.0. It is included on all seven Grok plans, the free one included. The catch is not the price but the meter: one weekly pool covers Chat, Imagine, Voice and Build together, and the vendor reports it only as a percentage.

Cheapest way in Free — you pay the model
Free tier Yes
Ways to buy it 7
Oldest check 2026-09-21
§ 01 / WHAT IT IS

What Grok Build actually is

Start with the name, because it is the most common way to land on the wrong page. SpaceXAI ships four things under the Grok brand, and two of them are agents with nearly identical names. Grok Bot is a platform of persistent agents living on a hosted cloud computer, billed by Cursor's parent company, and we cover it separately. Grok Build — this one — is a coding agent that runs on your own machine, in your own repository, from your own terminal. If you searched for a Grok coding agent, this is the product.

It installs with a shell one-liner or from npm, and starts by running grok inside a project. There is a full-screen mouse-interactive interface, a plan mode that blocks edits until you approve the plan, subagents that run in parallel with their own context windows and their own worktrees, and a headless mode for scripts and CI. Skills, hooks, MCP servers and an AGENTS.md convention file all work, and the default model is Grok 4.7, which replaced Grok 4.6 in that role on September 21, 2026. The harness and interface are open source in a public repository under Apache-2.0, which is why nothing on the pricing page sells you the client.

What distinguishes it commercially is not the tier list but the meter. Since June 2026 the vendor stopped giving each product its own daily limits and moved to one weekly pool shared across Chat, Imagine, Voice, the API and Build. That pool is shown as a percentage consumed, never as a quantity, and the vendor states that products draw on it unevenly — naming a long coding task as one of the expensive ones. Time with this agent is therefore spent from the same budget as your image generation, and neither the budget nor the exchange rate between them is published.

We have not run it. Authenticating requires a SpaceXAI account, and creating accounts is not something this desk does, so nothing below rests on our own session. Everything here was read on the vendor's own documentation, pricing matrix, licence and consumer FAQ on a single day, and the parts where two of those disagree are shown rather than resolved.

§ 02 / WHAT IT COSTS

What it costs, and what that buys

Full pricing →

Every tier includes the tool, so the table below is not a list of unlocks — it is a list of quantities, none of which the vendor states in a unit. The one route with a published rate is a key, and it carries a rule worth knowing before an agent runs on it.

Route Bought as Price
Grok Free Subscription $0/mo
SuperGrok Subscription $30/mo
SuperGrok Plus Subscription $100/mo
SuperGrok Heavy Subscription $300/mo
Grok Business Per seat $30/seat/mo
Grok Enterprise Per seat Contact sales
xAI API key API key At model rates
§ 03 / WHEN NOT TO

When the honest answer is use something else

You wanted the agents-on-a-cloud-computer product The mirror image of the warning on the other page. Grok Build is a terminal program on your machine that edits your working tree; it does not have its own computer, its own identity, or a schedule it wakes up on. If you want agents that sign into tools and run routines while you are asleep, that is Grok Bot, a different product on a different bill — and notably one that is not included on the free tier or on the business tiers, where this one is.
Your source cannot become training data by default On every individual plan, up to and including the most expensive one, the vendor's position is that it may use your content to train its models, and switching that off is a setting the user has to find. Worse for this specific product: the consumer FAQ that documents the opt-out names the mobile app and the website as the two places to do it, scopes itself to users of those two surfaces, and never mentions the command-line agent at all. The setting is account-level and most likely applies, but the vendor does not say so — and a commitment not to train arrives contractually only on the business tiers, where it is a row in the matrix rather than a checkbox.
You need to know your ceiling before you hit it There is no published allowance for any tier. The weekly pool is a percentage, it is shared with four other products that consume it at different rates, and the tiers are distinguished from one another by comparatives rather than numbers. If your objection to a tool is that you cannot forecast it, nothing on this page will fix that — a key billed per token is the only route here with arithmetic you can do in advance.
Your organisation is already locked down for Claude Code Grok Build deliberately reads Claude Code's managed settings file for permission rules, MCP allowlists and marketplace restrictions, which sounds like good news for a mixed fleet. The vendor then documents the gap in its own words: Claude Code's setting that disables bypass-permissions mode is not applied to Grok's always-approve, specifically so that Grok does not inherit a host's Claude Code lockdown. A fleet hardened through one tool is not hardened for this one until a separate policy file is deployed for it.
You expect the sandbox to be on, and to cover your keys Three defaults, all published by the vendor rather than discovered by us. The sandbox is off unless you turn it on, in a tool that executes shell commands. The built-in profiles do not permanently protect paths such as your SSH directory, and the documentation tells you to write your own deny list for them. And blocking network access for child processes is enforced on Linux only — on macOS it is a no-op for the two profiles whose names promise exactly that.
You need a price before you commit Procurement will struggle with two of the seven tiers. One is listed in the comparison matrix with a buy button and no price anywhere on the site, its link leading to a checkout that prices by the visitor's location. The other is the most expensive individual tier, whose figure is not on the pricing page at all but on the marketing page of a different product. Neither is unknowable, but neither is quotable from the page a buyer would naturally open.
§ 04 / HOW IT COMPARES

Nothing to set it beside, yet.

All comparisons →
!
No head-to-head covers this tool yet. The four Grok comparisons we publish are about the chat product and its subscription tiers, not the coding agent, so dropping one in here would describe a different thing that happens to share a brand. Pairs live in /compare/ and get written once rather than twice, which means this space stays empty until one genuinely covers Grok Build.
§ 05 / ALTERNATIVES

If this one is wrong for you

See the list →

The reasons people leave this tool are unusually specific — an unreadable meter, a training default, a sandbox that is off — and each points at a different replacement rather than at a general top ten.

Frequently asked.

What people search next.

Q · 01 Is Grok Build the same as Grok Bot? +
No, and they are close to opposites. Grok Build is a coding agent that runs in your terminal on your own machine, included on all seven Grok plans, open source, and billed by SpaceXAI. Grok Bot is a platform of persistent agents on a hosted cloud computer, included on only three tiers, closed, and billed by Anysphere — the company behind Cursor — through a Cursor account.
Q · 02 Is Grok Build open source? +
The harness and terminal interface are. The repository is public, written in Rust, and carries an Apache-2.0 licence file. One caveat for anyone whose compliance process reads package metadata rather than repositories: the npm wrapper declares Apache-2.0 but contains only a small bootstrap script, while the package that actually carries the compiled binary declares Proprietary. The model behind the agent is not open under any of these.
Q · 03 Which model does it use? +
Grok 4.7 by default since September 21, 2026 (it was Grok 4.6 before), which the vendor also recommends for coding on its API. A faster serving of the same model, Grok 4.7 Fast, is offered only here and in Cursor at twice the token rates, and is not included in the free tier. There is a separate and cheaper coding model in the catalogue as well. Beyond either, the client reads a user config file in which any model with an OpenAI-compatible endpoint can be declared and made the default, so it is not locked to this vendor's models.
Q · 04 Do I need a paid plan to use it? +
No. The free Grok plan carries Grok Build in the vendor's own comparison matrix, alongside every paid tier. This is genuinely unusual — most coding agents in this section are gated behind a subscription, and the sibling product from the same vendor is excluded from the free tier entirely. What the free plan does not come with is any statement of how much work it allows.
Q · 05 Can it run without a browser, on a server or in CI? +
Yes, and there are four documented ways to authenticate rather than one: a browser sign-in, a device-code flow for SSH sessions and containers, an external auth command for corporate identity providers, and a plain API key for unattended automation. Headless mode takes a prompt and emits structured output, with permission modes that either deny anything not explicitly allowed or auto-approve edits while still prompting for shell commands.
Q · 06 Does using it share my weekly usage with other Grok products? +
Yes. One pool covers Chat, Imagine, Voice, the API and Build together, and the usage screen reports a percentage with a per-product breakdown. The vendor says outright that different products cost different amounts of it and puts long coding tasks at the expensive end, so an afternoon of agent work and a batch of video generation compete for the same allowance.
§ 06 / EXPLORE

Keep reading.

Brand hub →
PRICING
Grok Build pricing
Seven tiers, one unreadable meter
↗ ALTERNATIVES
Grok Build alternatives
Sorted by why you are leaving
↗ TOOL HUB
Grok Bot, the other agent
Same brand, opposite shape
↗ TOKEN RATES
xAI per-token prices
What a key bills at
↗ RANKING
Coding agents, ranked
Computed from our own pairs
↗ SECTION
Every agent we price
Every tool, dated
↗
Y
Researched + verified by Yaroslav Vikhariev Founder · AI//COST · Verified Sep 21, 2026 · vendor docs, pricing matrix, repository licence, consumer FAQ
Methodology Report a correction More by Y.V.