Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Grok Build: SpaceX's $8-a-month coding agent

Written November 9, 2026

Grok Build 1.0.79 (Nov 8) is SpaceX's $8-a-month coding agent — 188 releases tracked, week-44 changes include vim editing and fail-closed hooks.

What it is

Grok Build is the terminal coding agent from SpaceXAI — the AI arm SpaceX formed after acquiring xAI in February 2026. It runs in your terminal, reads your repository, proposes a plan, edits files, runs commands, and splits large tasks across subagents in git worktrees. The harness is open source under the Apache 2.0 licence, in the Grok Build changelog and the xai-org/grok-build repository; the model itself is not open source. The default model is Grok 4.7, and the agent reads Claude Code and Cursor configuration files without conversion.

Version 1.0.79 shipped November 8, 2026 — the 188th Grok Build release tracked this year. Its changelog entry names no new features; the recent visible changes are the week-44 releases below, versions 1.0.71 to 1.0.73 from October 26 to 28.

What it costs

Grok Build comes with The X Subscription, which bundles Grok, the Cursor editor, and X Premium. The entry paid tier is Plus at $8 a month, then Premium at $30, Super at $100, and Ultra at $200, sharing one weekly usage pool across Grok and Cursor. There is also a limited free plan, and you can use Grok Build with a pay-as-you-go xAI API key instead: Grok 4.7 costs $2 per million input tokens and $6 per million output tokens. Prices are as reviewed on November 3, 2026; the vendor's pricing page is the authority before you pay.

What the week-44 releases changed

Experimental vim editing in the prompt (1.0.71). Set vim to all in the [keys] table of config.toml and the prompt gets Insert and Normal modes, with motions, operators, counts, and undo. The prompt's info line shows NORMAL or INSERT, and the cursor changes shape with the mode. It is off by default and labelled experimental, so its behaviour may still shift.

Hooks can fail closed (1.0.71). A hook now accepts a failClosed field. When a PreToolUse or UserPromptSubmit hook that sets it times out or crashes, the tool call is denied or the prompt is blocked, instead of going ahead. If you use a hook as a guard, this closes the gap where a crashed guard silently let things through. A fail-closed hook that crashes will block work, so check how reliable the hook is first.

Subagent output is clickable (1.0.71–1.0.73). Messages between two subagents appear as rows in the main transcript, and subagent and background task ids mentioned there open that subagent's view or the task's output on click. Relative file paths in agent output open on click when the file exists under the session's folder, and pull-request numbers open the pull request on the session repository's forge.

Smaller changes in the same three releases. A single Ctrl+C can quit when no draft or unfinished work would be lost (ui.single_ctrl_c_quit). New sessions can start in multiline mode, so Enter inserts a newline. The /delete confirmation now starts on Cancel, so pressing Enter no longer deletes the session. MCP tools built with FastMCP or the official MCP SDK send their result once instead of twice, using about half the context per call. Automatic updates stop retrying a new version that fails to start on the machine; each launch says to run grok update to see why.

What to know before pointing it at a repo

In July 2026, version 0.2.93 was shown uploading users' entire repositories — including git history and committed secrets — to SpaceXAI's cloud storage. SpaceXAI stopped the uploads with a server-side switch, said it would delete the data, turned default retention off, and published the source under the Apache 2.0 licence on July 15. No independent confirmation of the deletion has been published. Read the tool's data handling before you give it a repository that holds live secrets.

On capability, the published record is thin: a November 3 review notes that benchmarks so far do not show Grok beating Claude or GPT at coding. Treat Grok Build as a cheap way to try a full coding agent, not as a demonstrated upgrade over the ones you already use.

Verdict

At $8 a month inside a Plus subscription, Grok Build is the cheapest paid entry to a full coding agent — worth trying if you already pay for X or Cursor. Do not switch from Claude Code on price alone: nothing published shows it coding better, and the July upload incident is a reason to read its data handling first.

Steps

  1. Check what you already pay for. If you hold a Plus subscription or higher, Grok Build is already included. If not, Plus at $8 a month is the cheapest paid entry.
  2. Update, then try the week-44 changes. Turn on the vim prompt with vim = "all" under [keys] if you edit in vim, and set failClosed = true on any hook you rely on as a guard.
  3. Keep secrets out of the first repo you try. Use a repository without live credentials until you have read how the tool handles your code.
  4. Judge it on one real task. Run a task you have already done in your current agent and compare the result, the time it took, and the number of prompts — not the feature list.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Codex 0.167: run parallel sessions in one place · Model prices: what the major models cost · Copilot CLI 1.0.106: /code, /settings, GPT-6.1 Terra