AI ·
Grok Build
The coding agent in the terminal. Plan mode, subagents, skills, and the grok command. What it is for, and how I brief it.
Grok Build is xAI's coding agent. It was announced on 25 May 2026, an early beta for SuperGrok and X Premium+ subscribers. You install it, you open it in a repository, and it reads the tree. It edits files, runs the shell, searches the web, and reports back. The command is grok.
The install line from the docs:
curl -fsSL https://x.ai/cli/install.sh | bashWindows has a PowerShell equivalent from the same host. Then:
cd your-project
grokThe first launch opens a browser to sign in. A machine with no browser takes an API key in XAI_API_KEY. Scripts and CI get a headless form:
grok -p "Explain this codebase"Other editors can host the same agent through the Agent Client Protocol. On 15 July 2026 the harness was open-sourced, so the binary you install and the repository you can read are the same product from two sides. The product page is x.ai/cli. The launch post is Introducing Grok Build.
What the screen is for
The terminal UI is fullscreen, and it takes the mouse. The novelty of that wears off in a day. What stays is the set of controls around the model.
Plan mode. For a large change you ask for the plan and you make it stop. You approve, you comment on a step, or you throw the plan out. The edits come after the yes. This is the difference between an agent and a very fast colleague who has already touched forty files while you were reading the first paragraph.
Subagents. A large task splits. Each worker has its own context, and it can have its own git worktree, so two of them are not writing the same file under each other's feet. You remain the person who merges. Parallelism is a gift on research and on independent edits. It is a mess on one shared module, and the worktree is how the tool admits that.
Skills, hooks, MCP, and AGENTS.md. The agent reads the conventions already in the repo. A skill is a saved way of doing one kind of job, invoked by name or picked up when the task matches. A hook is a script that runs when a tool fires, which is where you put the formatter, the secret scanner, the thing that must happen every time. MCP is how it talks to a tracker, a database, a browser, a host you already pay for. AGENTS.md is the house rule, and it can sit at the root or deeper in a directory that has its own rules.
Three later additions changed the shape of a week with it. On 22 June 2026, /goal, for a long run that keeps going. On 23 July 2026, workflows: an orchestration script that fans a job out across many agents, checks the results, and reports once. On 16 September 2026, memory: notes written in the background and read back when you return to the project. Memory is convenient. It is also a place where a wrong assumption can survive the session that created it. Read it when a later session starts acting on a decision you do not remember making.
Which model
The harness and the model are separate knobs. Build launched beside a small coding model, Grok Build 0.1, on the API from 29 May 2026. Through the summer the default moved onto 4.5, then 4.6, then Grok 4.7. As of October 2026, 4.7 is the coding model the product points at, with a fast variant available in the agent. There is a model picker. Knowing which knob you turned saves an afternoon of blaming the prompt for a model change, or the model for a harness change.
On 19 August 2026 Grok Build arrived on the web and on mobile, on every plan. The phone is how you check a run. The terminal is still the desk where I read the diff.
How I brief it
I name the paths. I say what done looks like, usually a test, a command, or a page I can open. I say what to leave alone. I ask for the plan when the change is wider than I want to reread. Then I read the diff anyway.
The agent is fast at the boring middle of a change: the rename that touches twelve files, the test that follows a pattern already in the tree, the page that should look like its neighbors. It is confident when it is wrong, which is the same trait in every coding agent I have used. Confidence is a tone. The test result is the fact.
A good brief for this tool is the same shape as a good prompt anywhere else, with one extra line: the repository is the source, and files the task did not name are not an invitation. Give it a hand that matches the job. A shell that can do anything is the right hand for a migration, and a reckless one for a typo.
The companion that stays on after you close the laptop is Grok Bot. Build is for the repo in front of you. Bot is for the work that has to land in some other app, possibly while you are asleep. I keep those jobs apart on purpose.