We're excited to give you all early access to Grok Build. We built it as engineers who live in the terminal: people who reach for vim, think in keybindings, and treat the alt-screen as home. Our goal was to create something that feels native to that workflow giving you the best of grok and the terminal experience.
Here's a walkthrough of some of the features we are most excited about.
Image and video generation, right in the terminal
Grok Build ships with /imagine and /imagine-video commands that tap into our imagine models. Create assets for a webpage or creative project without leaving your session, and if your terminal supports the Kitty graphics protocol, preview them inline.
Full-transparency subagents
Most tools hide subagent work behind a title or collapse it entirely. We think you should see everything. Grok Build renders the full subagent trace in an overlay system layered on top of the main agent trace, with the same navigation controls. Nothing is hidden.
Keyboard and mouse friendly
The entire TUI is usable via keyboard alone, as any good TUI should be. But also supports mouse input and any mix of the two. Navigation takes cues from vim (j/k movement) while staying intuitive for mouse-first users.
Live command output in tool blocks
When the model runs a long-running terminal command, the output streams live into the tool block so both you and the model can follow along. You can also filter the output with regex or plain keywords, just like grep.
Faster worktrees
Worktrees let you cheaply spin up parallel workspaces on the same repository without cloning. We invested heavily in making them faster. On our monorepo, creation time dropped from ~1m 20s to ~20–30s on average. Subagents can use worktrees too, either by the main agent's decision or by your instruction.
We use this for our \/best-of-n\ approach which uses upto 5 subagents to accomplish a task and a judge agent to choose the best one (all subagents work in parallel on their own worktrees)
For technical details on how we optimized worktrees (including the overlayfs approach on Linux and CoW on macOS), see the linked post
Rewind to any point in the conversation
We built our own edit-tracking system, internally called the Hunk Tracker. That monitors changes at the hunk level, whether made by the agent or an external tool. The rewind flow shows you exactly which files can be cleanly reverted and which have been modified externally, so you're never guessing.
Plan mode with inline comments
The agent can ask you questions to iteratively build up a plan before writing code. Because we control the full alt-screen, plan mode is a first-class experience: you can leave inline comments on specific sections and interact with the plan naturally, not just approve or reject it wholesale.
/btw — side questions without breaking flow
Need to ask something unrelated while the agent is mid-task? Use /btw to fire off a side question without interrupting the main conversation. I use this constantly for quick tangential questions. And when I discover something new, Ctrl+Enter sends a live interjection to steer the model on the fly.
Import your settings from Claude Code
Your skills, MCPs, plugins, marketplace installs, and hooks are auto-detected and ready to use from the start. We wanted the transition to be seamless so you can skip setup and get straight to building.
What's next
This is v0. We're not done, not even close. Grok Build is the same tool engineers at SpaceXAI use every day, and we're building it to maximize real-world usefulness, not demos. We'd love your feedback on what's missing and what would make it better. We can't wait to see what you build with it.
