Grok Build

xAI has joined the terminal-coding-agent race, and Grok Build is its entry. Released in early beta in 2026 for SuperGrok and X Premium Plus subscribers, it is an extensible coding agent that runs three ways: as an interactive terminal UI, headlessly inside scripts and CI pipelines, or through the Agent Client Protocol so other applications can drive it. The category it is competing in — Claude Code, OpenCode, OpenAI’s Codex CLI — is already well-developed and well-loved, so the relevant question is not whether Grok Build is novel but whether it is good enough to belong in that company. On the evidence of the beta, it largely is.

The most strategically interesting thing about Grok Build is what it ships with on day one: native Agent Client Protocol (ACP) support, baked directly into the binary. ACP is an open standard that lets one agent talk to another without routing everything through the user’s chat window as a bridge. By building it in rather than bolting on an adapter, xAI is signalling that it sees Grok Build less as a standalone tool and more as a node in a larger mesh of cooperating agents. For teams building their own agent orchestration — bots that hand work to other bots — that out-of-the-box interoperability is a genuine differentiator, and it is the kind of forward bet that the more established tools are still catching up to.

On the developer-experience fundamentals, Grok Build covers the now-expected bases competently. Plan mode is the headline: before the agent touches your code, it proposes a plan, and you can approve it, comment on individual steps, or rewrite it entirely. Once approved, every change arrives as a clean diff for review rather than a black-box edit. For larger jobs, it delegates to specialised subagents — research, implementation, and review — running in parallel, each in an isolated git worktree so concurrent edits do not stomp on the main branch or each other. That worktree isolation is the same architectural idea that makes parallel agents safe in the best of the competition, and Grok Build implements it sensibly.

Adoption friction is mercifully low. Grok Build reads your existing AGENTS.md conventions, plugins, hooks, skills, and MCP servers out of the box, so pointing it at an established repo means it picks up your house style and tooling immediately rather than demanding a fresh setup. Headless mode (the -p flag) makes it straightforward to run inside automations and CI, and the same ACP support that enables agent-to-agent communication also lets you build your own orchestration on top of it. This is a tool clearly designed by people who have used the incumbents and copied their best ergonomics.

Pricing is where Grok Build diverges from the model many developers prefer. Rather than usage-based API billing or an open-source free tier, it is gated behind a SuperGrok subscription, with the higher Grok Build allotment carrying a real monthly cost during the introductory period. That is a meaningfully different bargain than OpenCode, the open-source agent PixlRun reviewed recently, which is free and lets you bring keys for 75-plus providers. It is also different from Claude Code, which is tied to Anthropic’s models but priced through familiar API and subscription paths. With Grok Build you are buying into xAI’s models and xAI’s subscription — no model-shopping, no BYOK escape hatch. Whether that is a fair trade depends entirely on how much you rate Grok’s coding ability, which is competitive but not the consensus leader.

That points to the honest caveats. This is an early beta, and it behaves like one — expect rough edges, occasional dead ends, and a feature set still settling. Because it runs xAI’s models exclusively, output quality rises and falls with Grok rather than with a model you chose, removing the flexibility that makes open agents resilient. And it is entering a field where the bar is high: Claude Code is an exceptionally polished single-vendor harness, and OpenCode is a free, model-agnostic juggernaut with millions of users. Grok Build has to be clearly better at something to pull developers away from tools they already trust, and right now its clearest “something” is the native ACP support and its tight fit for teams already living in the xAI ecosystem.

The verdict: Grok Build is a more-than-credible first showing — well-architected, fluent in the open agent standards before most rivals, and immediately comfortable in an existing repo. For developers already paying for SuperGrok, or for teams betting on ACP-based multi-agent orchestration, it is well worth a place in the rotation. The reasons to wait are the beta polish and the ecosystem lock-in: if you want a free, model-agnostic agent, OpenCode remains the pick, and if you want the most polished single-vendor experience, Claude Code still sets the standard. Grok Build does not yet displace either, but it has arrived close enough to matter, and the ACP bet may age very well.