Groundwork

SpaceXAI coding agent

Grok is the assistant; Grok Build is the agent.

Grok is SpaceXAI's assistant, and Grok Build is its terminal coding agent, powered by Grok 4.5 — a model pitched directly at frontier coding work. Groundwork installs the Grok Build CLI on personal machines; this page is the map.

When you'll want this You want to evaluate another frontier coding agent against the ones you already run, or your team or plan already includes Grok.

#Which Grok surface to use

Grok app & XThe assistant: chat, questions, and research in the Grok apps and inside X. The chatbot role GPT & ChatGPT describes — great at answering, can't see your project.
Grok Build CLIThe coding agent, in your terminal: an interactive TUI with plan mode, subagents that can run in parallel Git worktrees, or headless mode for scripts (grok -p "..."). Beta, bundled with the SuperGrok, SuperGrok Heavy, and X Premium+ plans.
Grok Build in other appsGrok Build speaks the Agent Client Protocol (ACP), so editors and apps that support ACP can drive it. Grok 4.5 is also selectable directly as a model inside Cursor, on all Cursor plans.
API & consoleGrok models by the token for your own tools and scripts, from the SpaceXAI console. Not available in the EU at launch (July 2026); SpaceXAI targeted EU access for later that month.

#The model underneath: Grok 4.5

On July 8, 2026, SpaceXAI released Grok 4.5, its first model built specifically for coding and agentic work, positioned as an Opus-class model at a markedly lower price ($2 per million input tokens, $6 per million output at launch; requests over 200K tokens of context bill at higher, tiered rates). It ships with a 500,000-token context window — smaller than the 1M window on Grok 4.3. On SpaceXAI's own SWE-Bench Pro results, Grok 4.5 resolves 64.7% of tasks while using roughly 16K output tokens per task on average, well below heavier models on the same benchmark — the token-efficiency story behind the low API price. It powers Grok Build, is selectable in Cursor, and is available from the API console. The durable takeaway isn't the launch ranking — it's that the frontier now has another serious coding agent, and the evaluation loop below is how you decide if it earns a place in your rotation.

#Trying Grok Build

Groundwork installs the CLI on personal machines (the grok-build Homebrew cask, which puts the grok binary on your path), so trying it is one command in a project folder:

grok
$ grok

It reads AGENTS.md out of the box, along with your skills, hooks, and MCP servers — which is exactly why Groundwork keeps its agent contract in that tool-neutral file (see AI agents & rules). Evaluating a new agent is the same loop as directing a familiar one: start it in a clean repo, ask for inspection before edits, approve narrowly, and judge it by the diff it leaves. If you can run that loop with Claude Code or Codex, you can run it here — that transferability is the point.

#Daily commands

grokOpen the interactive TUI in the current project.
grok -p "explain this repo"Run headlessly with a single prompt, for scripts, bots, or CI.
grok inspectShow what Grok discovered in the project: config sources, instructions, skills, plugins, hooks, and MCP servers.
grok --resume <session-id> / grok -cReopen a specific past session, or continue the most recent one.
/plan <description>Enter plan mode: draft an approach without editing files, then review it before any code changes.
/model <name>Switch the active model mid-session.
/context & /usageThe two gauges: context-window usage and credit usage. Check both before long work — see Know your AI budget.
/compactCompact the conversation history to free up the context window.
The binary is installed; the plan is yours

Groundwork installs the tool, not the account: SpaceXAI has offered a free trial, and ongoing use runs on a SuperGrok or X Premium+ subscription or API credits — check the Grok Build page for what access costs today. Work machines skip it along with the rest of the personal AI apps; see Platforms.

#Official references

  • Introducing Grok 4.5 — for anyone deciding whether the model is worth evaluating at all: the launch announcement, pricing, and positioning.
  • Grok Build beta — for a reader ready to install: the coding agent's home page, the plans it ships in, and the install script.
  • Grok Build docs — for daily use: CLI, ACP, config, and permissions reference.
  • Modes & commands — for the full command surface: every slash command, plan mode, and session management in one page.
← PrevGLM & open models Next →Everyday workflow