Grok Build is xAI's AI coding agent: it plans, writes and tests code from your terminal, and its Build mode on grok.com does the same work inside a chat window. The agent's source code sits on GitHub under an Apache 2.0 license, a rarity in this market. Individual access runs through a SuperGrok subscription at $30 a month, while the grok.com Build mode stays limited to SuperGrok Heavy during its beta. You describe a project, approve the plan, and a publishable link comes out the other end.
Grok Build lives first in a fullscreen command-line interface, mouse support included. For a tricky job you start in Plan Mode: the agent drafts its roadmap and every file edit stays blocked until you sign off. You can comment on a single step, rewrite it, or wave the whole thing through.
Once the plan clears, delegation kicks in. Up to 8 subagents work in parallel, each with its own context window, down to separate Git worktrees. The rest works out of the box: AGENTS.md files, plugins, hooks and MCP servers, plus a memory that keeps your project conventions between sessions and commands like /imagine or /loop for generated images and repeated checks. It slots naturally into the wider family of autonomous AI agents, with a keyboard-first feel.
| Interface | What happens there | Access |
|---|---|---|
| Fullscreen TUI | Live session with visible plan and subagents | SuperGrok subscription |
| Headless CLI | Scripts, bots, automations | Subscription or API key |
| ACP mode | Integration inside other apps | Subscription or API key |
| Build mode (grok.com) | Apps, games and dashboards published to a link | SuperGrok Heavy (beta) |
| grok-build-0.1 API | Your own agent loop | Pay per token |
Build mode on grok.com produces a working application from a single message and previews it right inside the chat, on the web or on iOS and Android. You type "build a 16-step drum sequencer in my browser" and the thing starts playing in front of you (the code stays one tab away, though nothing forces you to look at it, which is rather the point). Edits happen in plain language: layout, colors, an extra feature.
Publishing takes one button: a grok.me link hosted by xAI or a domain you already own, with GitHub export if you want to keep iterating elsewhere. Working tools with real state and logic, arcade games and 3D worlds, dashboards fed by connectors, personal sites: the launch examples cover all four.
On the Artificial Analysis coding agent index, Grok Build running Grok 4.5 scored 76 points, a single point behind Claude Code, while completing the benchmark tasks at a fraction of the cost. At launch the agent posted 70.8% on SWE-bench Verified. The real contest, then, is less about raw scores and mostly about what each finished task costs you.
Money-wise, individual access goes through SuperGrok at $30 a month, where the original beta demanded the $300 Heavy tier. API developers can pull the grok-build-0.1 model at $1 per million input tokens and $2 output, served above 100 tokens per second. xAI adjusts these numbers often, so treat the official pricing page as the last word.
No. The CLI works with a SuperGrok subscription billed at $30 a month, or with an API key paid per token. xAI has occasionally opened free trials of the model inside the CLI. The Build mode on grok.com, meanwhile, requires the SuperGrok Heavy tier for as long as the beta lasts.
Benchmarks put them one point apart on the Artificial Analysis index (76 versus 77), yet Grok Build finishes the same tasks for far less money and its source code can be audited. Claude Code keeps the maturity edge and a larger ecosystem; on a tight budget or heavy workloads, Grok Build holds its ground.
Yes. xAI published the project in the xai-org/grok-build repository under an Apache 2.0 license, answering questions about repository syncing and code privacy. That means you can inspect what the agent does, fork it, or even point it at a custom model through its configuration files.
It does. Installation goes through PowerShell on Windows and a one-line script on macOS and Linux. The CLI then works with any repository and any language, with browser-based sign-in by default and API keys for machines without a graphical environment, such as servers and CI runners.
Verdict: From a patched shell script to a 3D game behind a shareable link, the range is wide: terminal regulars, automation-minded teams and first-time builders on the chat side will each find their footing here.
