Grok Build opens beta with /loop, /imagine, and best-of-n agent workflows

• xAI opened an early beta of Grok Build for SuperGrok Heavy subscribers, and xAI's launch post frames it as a CLI for coding, app building, and workflow automation.
• Early screenshots from Everlier's feature roundup show slash-command workflows including /loop, /btw, /implement, /review, plus built-in /imagine and /imagine-video generation.
• The most interesting systems leak through Everlier's screenshots and Everlier's pricing card screenshot: parallel subagents in isolated worktrees, a /best-of-n tournament mode, self-validation with --check, and a memory flow described as "dreaming."
• xAI's own promo image, amplified by thekitze's repost, pitches a "fast and flicker-free CLI," while Everlier's screenshot of the offer says Grok Build access sits inside a discounted $99 per month SuperGrok Heavy intro offer that later returns to $300.

You can open the beta landing page, scan the launch post, and then jump straight to Everlier's screenshots, which expose a surprisingly opinionated command set. There is even a documented /best-of-n tournament flow, plus image generation already wired into the CLI.

xAI's announcement is short: Grok Build is an early beta for SuperGrok Heavy users, and the product pitch is a single surface for coding, building apps, and automating workflows. The install flow in thekitze's repost points to curl -fsSL https://x.ai/cli/install.sh | bash, which makes this look more like a Claude Code or Codex-style terminal product than another browser agent.

The pricing card in Everlier's screenshot of the offer adds the rest of the launch framing:

• Exclusive access to Grok Build
• Flicker-free CLI
• Parallel subagents and worktrees
• Smarter planning and self-review
• MCP, skills, plugins, and hooks
• Headless mode for scripting and CI/CD

The early command list from Everlier's feature roundup is unusually broad for a coding CLI. It mixes agent control, implementation workflows, and native media generation in one prompt surface.

• /loop
• /btw
• /implement
• /review
• /imagine
• /imagine-video
• /compact
• /compact-mode
• /theme

The media angle is not theoretical. The command picker in Everlier's screenshot of /imagine shows text-to-image and text-to-video sitting directly inside the terminal, and Grok's image stack already has creators publishing outputs such as gizakdag's playing-card demo and carolletta's Heartbreak video made in Imagine Agent Mode.

The clearest workflow reveal is /best-of-n. According to Everlier's /best-of-n screenshot, the CLI can spawn 2 to 10 subagents in parallel, put each one in its own isolated git worktree, score the results on correctness, code quality, and safety, then apply the winning changes back to the main workspace.

That same screenshot makes the flow concrete:

1. Spawn N subagents in parallel.
2. Let each candidate implement the task independently.
3. Compare outputs with an orchestrator.
4. Apply the winner to the main workspace.
5. Show a comparison table and winner index.

A separate screenshot in Everlier's verification panel shows python3 -m py_compile, a live npm API test, and a post-run rebuild command, which lines up with Everlier's claim that --check makes the agent self-validate completions in headless mode.

The same batch of screenshots exposes a few smaller design choices that say a lot about how xAI wants this tool used. Everlier's compact-mode and theme screenshot shows a 512K token context meter, compact mode toggles, timestamp toggles, and named themes including groknight, grokday, tokyonight, and rosepine-moon.

Everlier also says the memory system uses "dreaming," a consolidation step for collected memories Everlier's feature roundup. That language makes Grok Build sound less like a raw shell wrapper and more like an agent harness with long-running state.

The beta pitch lands in a moment when terminal-first agent setups are already sprawling across tmux, Warp, Zed, and helper models. Everlier's workspace screenshot shows exactly that kind of environment, with separate project sessions and multiple agent tools running side by side.

That context makes the product positioning in Everlier's pricing card screenshot more legible. Grok Build is not being sold as a chat upgrade. It is being sold as another seat in the emerging stack of always-on coding agents, with worktrees, hooks, headless runs, and parallel delegates.

Access is narrow for now. xAI's launch post says the beta is limited to SuperGrok Heavy subscribers, while Everlier's pricing card screenshot shows a six-month promo at $99 per month, marked as 67 percent off, before returning to $300 per month.

That same card also confirms two more launch details that did not make the main announcement copy: headless mode is part of the package, and extensibility is supposed to include MCP, skills, plugins, and hooks.