Grok Build

Overview

Status note (August 2026): Two things changed since the June revision of this profile. First, the company: SpaceX absorbed xAI in a February 2026 all-stock deal and completed the rebrand to SpaceXAI on July 6, 2026 — the Grok product brand is unchanged. Second, and more consequentially: on July 15, 2026, SpaceXAI open-sourced the entire CLI under Apache 2.0 at github.com/xai-org/grok-build — roughly 72 hours after a security researcher published wire-level evidence that Grok Build had been silently uploading users' complete Git repositories to a SpaceXAI-controlled cloud bucket. This profile has been updated against a direct review of the public source.

Grok Build is SpaceXAI's terminal-native coding agent — a fullscreen, mouse-interactive TUI written in Rust, installed via curl -fsSL https://x.ai/cli/install.sh | bash, that plans, edits, and runs code from the terminal with diff-based review, runs headlessly for CI, and embeds in editors via the Agent Client Protocol (ACP). It launched in early beta on May 14, 2026 as xAI's answer to Claude Code and Codex CLI, arriving roughly a year behind both.

Key stats (as of August 2026): grok-build-0.1 model, 256K token context, $1.00/$2.00 per million input/output tokens via API; client open-sourced under Apache 2.0 — 844,530 lines of Rust, ~24K GitHub stars within three weeks.

---

Open Source, Under Duress

The open-sourcing is the defining event of Grok Build's short life, and the sequence matters:

• ~July 12–13: Researcher cereblab published mitmproxy captures showing Grok Build 0.2.93 transmitting full Git bundles — every tracked file plus complete commit history, committed secrets included — to a Google Cloud Storage bucket named grok-code-session-traces. The "improve the model" privacy toggle did not stop the uploads; a server-side flag eventually did.
• July 15: SpaceXAI open-sourced the client, said uploaded data was deleted and retention disabled by default, and reset usage limits for all users.

What the repo actually shows, from direct review:

• 844,530 lines of Rust (only ~3% vendored), published as a single initial commit with no development history, then synced one-way from the internal monorepo — a SOURCE_REV file records the monorepo SHA.
• Open for audit, not collaboration. External contributions are explicitly not accepted. This is source-available trust repair in Apache 2.0 clothing, and on those terms it works: anyone can now verify what the tool sends home.
• A documented fast-follow lineage. The third-party notices disclose in-tree source ports of tool implementations from openai/codex and sst/opencode, with Apache §4(b) change notices — the tool layer literally began as ports of its competitors.
• The model stays closed. grok-build-0.1 remains API-gated behind SpaceXAI subscriptions. You can read and modify the client; it still calls a proprietary model over the network.

---

What Shipped vs. the Rumors

---

Features

Plan Mode

For complex tasks, Grok Build starts in plan mode: it drafts a structured, step-by-step approach that the developer can approve, comment on per-step, or rewrite entirely before execution begins. Once approved, every change surfaces as a clean reviewable diff — the same review pattern Claude Code and Codex established.

Subagents and Worktrees

Larger tasks are delegated to specialized subagents that run in parallel — research, implementation, and review concurrently. Subagents launch in their own isolated git worktrees, preventing conflicting edits on the main branch.

Headless CI Mode

grok -p "task" runs non-interactively with optional streaming-json output. Default permission mode is ask (prompt per tool call); --always-approve enables fully automated pipelines.

Extensibility

MCP servers, AGENTS.md conventions, plugins, skills, and hooks are supported, with grok inspect exposing configuration sources — a near-exact mirror of the Claude Code extension surface. The shipped user guide also documents sandboxing, theming, and slash commands.

Editor Embedding (ACP)

The agent runtime speaks the Agent Client Protocol, so the same engine that drives the TUI can be embedded in editors — closing part of the "terminal-only" gap flagged at launch without SpaceXAI building an IDE plugin per editor.

---

Model and Pricing

---

Viability and Adoption

Grok Build is two years into the company's pattern of uneven developer-tool follow-through, and the trajectory since launch has been whiplash: shipped in May, caught exfiltrating repositories in mid-July, open-sourced 72 hours later. The repo drew roughly 24K GitHub stars in its first three weeks — real attention, though stars measure curiosity about 844K lines of frontier-lab agent code as much as they measure users. SpaceXAI still has published no SWE-Bench Verified score for grok-build-0.1; the older 70.8% figure belonged to the now-deprecated grok-code-fast-1 and trails the mid-2026 leaderboard (Claude Opus 4.7 Adaptive at 87.6%, OpenAI Codex at 85%) by roughly 17 points. No verified user or revenue numbers for Grok Build have been published as of August 2026.

Cautions

• The July privacy incident is disqualifying history for some buyers. Before July 13, 2026, the CLI uploaded full Git bundles — tracked files, complete commit history, committed secrets — regardless of the privacy toggle. Anyone who ran Grok Build in a Git repository before that date should treat every credential in tracked files and commit history as potentially transmitted, including keys committed and later removed. SpaceXAI says the data was deleted and retention is now off by default, but the fix that mattered was a server-side flag, not the user-facing control.
• Open source ≠ open development. The repo takes no external contributions and is synced one-way from the monorepo; audits are possible, influence is not.
• Pricing is the category's steepest. Full access launched behind SuperGrok Heavy at roughly $300/month versus Claude Code or Codex CLI bundled into $20/month plans — independent reviewers called out the ~15x cost differential as the product's biggest adoption barrier.
• Benchmark claims rest on SpaceXAI's own evaluation suite, without independent validation at launch.
• The hype gap is documented. Critical coverage found launch-cycle claims about context window, Plan Mode defaults, and the model's identity to be overstated or wrong — a caution for anyone evaluating the tool from secondhand coverage.
• Youngest track record in the category. Every competing foundation-lab agent has 12+ months of production hardening; Grok Build has months, one of which included the repo-upload incident.

---

Competitive Position

Strengths: Fully auditable Apache 2.0 client — 844K lines of Rust anyone can read, a transparency bar Claude Code does not meet. Genuine parallel subagent architecture with worktree isolation. Cheapest coding-model API pricing in the lineup ($1/$2 per M tokens). Full extension stack (MCP, AGENTS.md, skills, hooks, ACP) at launch rather than added later.

Weaknesses: Trust deficit from the July repo-upload incident. Most expensive subscription gate in the category. Smallest context (256K vs 1M for Claude Code and Gemini CLI). No GitHub-native integration. No published independent benchmarks. SpaceXAI model lock-in — the open client only talks to the closed model.

---

Bottom Line

Recommended for: Existing SuperGrok / X Premium Plus subscribers who want a Claude Code-style agent within the SpaceXAI ecosystem; teams that value an auditable client above all and have verified the current build's network behavior themselves; researchers studying how a frontier lab structures an agent harness — the public source is the most complete look inside any big-lab coding agent to date.

Not recommended for: Anyone whose codebase ran through Grok Build before July 13, 2026 without a subsequent credential rotation; security-sensitive teams for whom the incident is disqualifying regardless of remediation; anyone optimizing for cost (15x the entry price of Claude Code/Codex); buyers who require independently benchmarked models.

Outlook: The open-sourcing was crisis response that doubled as genuinely useful transparency — ~24K stars in three weeks says the audit story landed. But the structural read is unchanged: this is a fast follow (the tool layer literally ports Codex and opencode code), the model is closed, development is closed, and the trust repair has to hold through the next incident. Watch whether SpaceXAI publishes independent benchmarks for grok-build-0.1, drops the subscription gate toward a $20-30 tier, keeps the public repo synced as promised, and whether the ACP surface turns into real editor adoption.

---

Research by Ry Walker Research