Grok Build

Grok Build is xAI's entry in the terminal-coding-agent race — a CLI that plans, edits files, runs shell commands, and coordinates subagents from the command line, powered by the Grok models behind the xAI API. It shipped as an early beta in May 2026 and sits, post the SpaceX/Cursor deal, inside a company that suddenly owns both the model and the dominant editor. This page is the orient-and-decide surface — xAI's docs and pricing page own the exact model IDs, tiers, and current prices, all of which are moving fast on a beta product.

What it is

A terminal-native agentic coding tool: describe the change in natural language, it produces a plan, and — on approval — edits files, runs commands, and manages dependencies across a repo. Installed with a single curl | bash, it's clearly modeled on the Claude Code / Aider lineage of "the agent lives in your terminal, not a separate IDE."

Its one genuinely distinctive design choice is parallelism: rather than one agent working one thread, Grok Build fans a task out to multiple subagents — up to eight — each in its own git worktree, so independent workstreams run concurrently without stepping on each other's files. That worktree-per-agent model is the same isolation discipline a careful operator already uses by hand; Grok Build makes it the default execution model.

When to use it

Reach for it when:

• You want to evaluate parallel-agent coding — several independent changes advancing at once, each sandboxed in its own worktree — which is the one thing here that isn't just a Claude Code re-run.
• You're already paying for xAI's top subscription tier for other reasons and the CLI comes bundled — the marginal cost to try it is then zero.
• You need a headless coding agent in CI — the -p non-interactive flag is built for pipelines, and MCP/ACP support lets it reach your existing tools.
• You're deliberately betting on the SpaceXAI stack — model (Grok), agent (Grok Build), and editor (Cursor) under one roof — and want the first-party agent.

Skip it when:

• You already run Claude Code on a flat-rate Max subscription and it works — Grok Build is a lateral move behind a different paywall, not an upgrade (see Claude Code).
• You want $0 and fully local — this is a paid, gated, closed-cloud agent; the free/local answer is Aider on Ollama.
• The prompt carries private operating data — same rule as any cloud coding agent: it ships your code to a third party tied to X.
• You need stability — it's an early beta on a fast-release cadence; APIs, model names, and pricing will move under you.

At a glance

Exact model IDs, the context window, tier names, and current prices live on xAI's docs and pricing page — treat those as canonical over anything quoted here; a beta product's numbers drift week to week.

Where it fits on the GL stack

Framed conditionally — watching, not adopted.

• The one idea worth stealing is the parallel-worktree model. Fanning a task across subagents that each own a git worktree is exactly the isolation discipline a careful operator already applies by hand — seeing it as a first-class execution mode is the interesting part, whichever agent ends up running it.
• It's a competitor to the daily driver, not a complement. The working setup here is Claude Code on a flat Max subscription. Grok Build asks you to pay a second premium subscription to a closed cloud model to do the same job — the bar for switching is "meaningfully better," and an early beta doesn't clear it yet.
• Its real significance is the stack, not the CLI. After the Cursor acquisition, SpaceXAI holds model + agent + editor. Grok Build is the terminal expression of that; the thing to watch is whether the integrated stack compounds into something the piecemeal competition can't match — which is a strategic question, not a "install it today" one.

Gotchas

• Gated behind a premium subscription. This isn't a free CLI you point at any key — access rides on xAI's top paid tiers. Budget the subscription, not just the tokens.
• Beta cadence. Early-beta means fast, breaking change — install command, flags, model names, and prices will move. Don't wire it into anything load-bearing yet.
• "Parallel agents" is a coordination problem, not just a feature. Eight subagents editing eight worktrees is powerful and easy to lose track of — the review surface (which agent changed what, and why) matters more here than in a single-threaded tool.

Risks

• Data leaves the machine. Every run ships your code to a closed cloud model tied to X — the headline risk for a local-first posture, not a footnote. Gate hard on what's allowed over the wire.
• Second-subscription creep. Adopting it means paying for xAI's premium tier on top of an existing Claude Max subscription — two paywalls to do one job. That needs a reason beyond curiosity.
• Vendor concentration. Betting on the SpaceXAI stack for model + agent + editor is convenient and consolidating — and it's a lot of eggs in one fast-moving vendor's basket.

Related

• xAI Grok API — the model API Grok Build is built on; the same OpenAI/Anthropic-compatible backend, wrapped in an agent.
• Claude Code — the incumbent terminal coding agent it's chasing, and the flat-rate daily driver it has to beat.
• Cursor — the editor half of the post-acquisition SpaceXAI coding stack.
• Aider — the open-source, $0, local-Ollama alternative when a cloud agent is the wrong call.