Grok Build

Learning Objectives

• Understand what Grok Build does and where it sits among terminal coding agents
• Evaluate the July 2026 open-source release in the context of the security finding that preceded it
• Distinguish source transparency from community-governed open source

What Is Grok Build?

Grok Build is the terminal coding agent from SpaceXAI — the company formerly called xAI, operating under SpaceX since the February 2026 acquisition. It runs as a full-screen interface in your terminal that understands a codebase, edits files, executes shell commands, searches the web, and manages long-running tasks. It is built around SpaceXAI's own Grok models, which is the pairing the product is really designed for: a model tuned hard for agentic coding, driving an agent harness written to exploit that.

On July 15, 2026, SpaceXAI published the agent's complete Rust source on GitHub under an Apache 2.0 license — roughly 844,000 lines covering the agent runtime, the terminal interface, the tool layer, workspace and version-control handling, sandboxing, and Model Context Protocol integration.

Source Transparency Is Not Community Open Source

The license is Apache 2.0, but the project is not run as a community effort. External contributions are not accepted, GitHub issues are disabled, and the repository is described as a periodic sync from SpaceXAI's internal monorepo rather than a live development branch.

Key Capabilities

A Terminal-Native Agent

Grok Build lives where the work happens rather than in a separate editor window. The interface is full-screen and mouse-interactive, with scrollback and modals, and it drives the same primitives a developer would: reading files, editing them, running commands, and checking results.

Which Grok Model It Runs

Grok Build ran on Grok 4.5 in July 2026, moved to Grok 4.6 on August 12, 2026, and has offered Grok 4.7 since its September 21 release. SpaceXAI's Grok Build product page still names Grok 4.6 as the model powering it, so check the model shown in your session. All three generations cost the same two dollars per million input tokens and six dollars per million output, which is the product's clearest economic argument: an agent loops many times per task, so the model bill compounds, and SpaceXAI has held that rate while the models improved. Grok 4.7 roughly doubled Grok 4.6's Terminal-Bench score on SpaceXAI's own numbers, though it still trails Claude Fable 5.1 by about twenty points there.

Written in Rust

The agent is roughly 99.6% Rust. For a tool that spawns processes, touches the filesystem, and runs in a developer's shell, the choice buys startup speed and memory safety — relevant given that the agent's job is to execute code on your machine.

Grok Build vs. Other Terminal Coding Agents

Pricing

The agent code is free; the model calls are not. Because Grok Build is an agent that loops, the model bill — not the license — is what a team actually pays for.

Related Tools

• Grok 4.7 — SpaceXAI's current flagship, available in Grok Build since September 21, 2026
• Grok 4.5 — the model Grok Build ran on in July 2026
• Claude Code — Anthropic's terminal coding agent, closed source
• OpenAI Codex — OpenAI's coding agent, closed source
• Cursor — the SpaceX-owned editor Grok 4.5 was trained alongside
• Gemini CLI — Google's Apache 2.0 terminal agent, which did accept contributions, discontinued in June 2026

Strengths

• Auditable — the full Rust source is published, so its data handling can be inspected rather than trusted
• Stable model pricing — three Grok generations at the same two dollars and six dollars per million tokens, which matters for an agent that loops
• Terminal-native — full-screen interface with scrollback, modals, and mouse interaction, without leaving the shell
• Fast and memory-safe — a Rust implementation for a tool that spawns processes and touches the filesystem
• Forkable — Apache 2.0 permits modification and redistribution, including a privately patched build

Limitations and Considerations

• The security finding is the context — the agent was reported uploading complete Git repositories, commit histories included; audit what your build sends before pointing it at a repository with sensitive history
• Not community-governed — contributions refused, issues disabled, synced periodically from an internal monorepo; you can fork it but not steer it
• Vendor-reported model benchmarks — the Grok models' benchmark numbers come from SpaceXAI's own charts
• Trails on terminal tasks — even Grok 4.7 scores well below Claude Fable 5.1 on Terminal-Bench, the benchmark closest to what a terminal agent does
• Single-model gravity — built around SpaceXAI's own models, so the economics that make it attractive are tied to SpaceXAI's pricing
• Young ecosystem — SpaceXAI's developer tooling and support remain less mature than OpenAI's, Anthropic's, or Google's

Key Takeaways

• Grok Build is SpaceXAI's terminal coding agent — a full-screen interface that reads code, edits files, runs commands, and manages long tasks, on Grok 4.6 by default with Grok 4.7 available since September 21, 2026
• On July 15, 2026 the company published roughly 844,000 lines of Rust under Apache 2.0, covering the runtime, interface, tool layer, sandboxing, and Model Context Protocol integration
• The release followed security research showing the agent had been uploading developers' entire Git repositories, commit histories and credentials included
• It is source transparency rather than community open source: contributions are refused and issues are disabled, so you get auditability and the right to fork, but no say in direction
• The agent is free; the model calls are not — and for a looping agent, the model bill is the real cost