Grok Build

What it is

Grok Build is xAI's coding harness. The harness and its terminal interface were open-sourced on 16 Jul 2026, while the hosted service around them stays proprietary — so you can read the loop, but the thing most people use is still xAI's.

Since 19 Aug 2026 it has been included with every xAI plan. That is a more aggressive distribution decision than it first sounds: the web and mobile tiers get the coding agent too, not just the developer-facing subscriptions. Whatever else is true of it, Grok Build is in front of a lot of people who did not go looking for a coding agent.

How the loop works

Three surfaces — terminal, web and mobile — onto the same agent. The terminal CLI is the one to use for real work; the web and mobile clients matter mostly because they exist at all, and because they make starting a job from a phone a normal thing rather than a novelty.

The feature worth attention is session memory, generally available since 16 Sep 2026. It is explicit rather than ambient: /memory to manage what is retained, /dream to consolidate, and lock and unlock controls over what can change. Most harnesses either forget everything between sessions or accumulate context silently. Grok Build gives you handles on it, which is the right shape even if the commands take some learning.

Models are the Grok line only. Grok 4.6 is the default at $2 per million input tokens and $6 per million output. A harness-specific model, grok-build-0.1, arrived on 8 Sep 2026 at $1/$2 with a 500k context window — cheaper and longer-context, aimed at the agent workload rather than general chat.

Isolation and permissions

Split between the hosted service and the local CLI. Work done through the hosted side runs on xAI's infrastructure; the CLI runs on your machine with your permissions, like any terminal agent. There is no distinctive sandbox story here, so treat the local CLI the way you would treat the others: contain it yourself if the task warrants it.

Who it is for

High-volume users on an xAI subscription. At $2/$6 for Grok 4.6, and $1/$2 for grok-build-0.1, the economics are the main argument — if you are running an agent for hours a day, the per-token difference against frontier pricing compounds quickly. The all-plans inclusion also means there is nothing extra to buy to try it.

Limits

Single-vendor by construction: the Grok line or nothing. If you want to switch models when a better one appears, this is the wrong tool. The open-sourced harness softens that a little — the code is readable — but the service you are paying for is not portable.

It is also moving very fast. The CLI shipped daily updates through September 2026 (v1.0.34 on 19 Sep, for instance), which is good for fixes and bad for stability. Pin a version if you depend on behaviour staying put. And the cheaper grok-build-0.1 is early — version 0.1, released 8 Sep 2026 — so do not assume it matches Grok 4.6 on hard work just because it is priced for the same job.

Alternatives on this board

• Claude Code — the other single-vendor terminal agent, further along on permissions.
• Codex CLI — open source and sandboxed by default.
• OpenCode — provider-agnostic, if avoiding model lock-in is the priority.

Sources

• x.ai — product announcements and plan inclusion
• xAI documentation — model pricing and context windows; the open-sourced harness repository is linked from there