Grok Build

Grok Build is an agentic coding tool and command-line interface (CLI) developed by xAI, the artificial intelligence company founded by Elon Musk. Launched in early beta in mid-May 2026, it is xAI's first dedicated coding agent, a category of AI code generation software that runs in a developer's terminal, reads and edits a codebase, executes commands, and carries out multi-step engineering tasks described in natural language.^[1][16] At launch, Grok Build was built around a purpose-trained model called grok-build-0.1 and positioned xAI directly against the two established leaders in the space, Anthropic's Claude Code and the OpenAI Codex CLI. Coverage at launch widely framed it as a catch-up move: Musk had publicly conceded that xAI trailed rivals on coding, and the product shipped roughly a year after Claude Code and Codex reached general developers.^[4][6][7][8]

Overview

Grok Build is distinct from earlier xAI coding efforts. It is not a chat feature inside the Grok assistant, nor is it the same thing as Grok Code Fast 1, the lower-cost coding model xAI released in 2025. Instead, Grok Build is a standalone command-line program (invoked simply as grok). Its original release wrapped a new model, grok-build-0.1, which xAI's documentation describes as the successor to grok-code-fast-1.^[2] The product therefore has two separable layers: the agent and CLI ("Grok Build") and the underlying large language model used for inference. The launch configuration used grok-build-0.1, while SpaceXAI made Grok 4.5 the default model on July 16, 2026; current versions can also use custom model endpoints.^[2][26]

xAI first teased the tool in April 2026, when Musk predicted it would approach the coding ability of a frontier model by May and match or exceed it by June. The early beta opened in mid-May 2026, with most press coverage dating the announcement to May 15, 2026.^[9][10] Access was initially restricted to the top SuperGrok Heavy subscription tier. About a week and a half later, around May 24 to 25, 2026, xAI broadened availability to all SuperGrok and X Premium+ subscribers, sharply lowering the price of entry.^[11]

The launch landed during a turbulent period for the company. Following xAI's early-2026 combination with SpaceX, reporting noted significant staff departures, even as the product itself was described as technically credible. Grok Build is part of a broader 2026 xAI developer push that also included additional tooling and connectors aimed at software teams.

Capabilities

Grok Build is an agentic CLI: a developer points it at a project directory, describes a task in plain English, and the agent inspects the repository, locates the relevant files, and proposes and applies changes. Its headline architectural feature is parallelism. Grok Build can spawn up to eight subagents that run simultaneously, and a distinctive design choice is that each subagent operates in its own Git worktree.^[12] This isolates speculative work on separate branches so that competing approaches can be explored and later merged, rather than all agents mutating a single shared workspace.

Work flows through a plan-first model often summarized as plan, search, and build.^[1] The agent drafts an execution plan that the developer can review and approve before any code is written, then searches the codebase to ground its changes, then carries out the edits. This emphasis on a reviewable plan reflects xAI's stated focus on developer trust and oversight rather than fully opaque automation.

The tool ships in several interface modes. There is an interactive terminal UI for hands-on sessions; a headless mode (using a -p prompt flag) suitable for scripting and continuous-integration pipelines; and integration into other applications through the Agent Client Protocol (ACP).^[12] The CLI supports switching models in session (for example with a /model command) and inspecting the active model, and it is designed to run locally in the terminal and operate directly on the developer's local repository. In the original hosted configuration, model inference was served through xAI's API.^[2] Installation is a single shell command on macOS and Linux or a PowerShell command on Windows, with authentication handled through the browser on first launch or via an API key in non-interactive environments.^[2]

xAI has also described an "Arena Mode," an automated evaluation layer that would run multiple agents against the same problem, score their outputs, and rank them before a developer reviews the result.^[1] As of the early beta, multiple reviews indicated Arena Mode was on the roadmap rather than fully live, so it should be treated as an announced capability rather than a shipped one.

Open-source release and architecture

On July 15, 2026, SpaceXAI released the source code for Grok Build's coding-agent harness and terminal UI in the public xai-org/grok-build repository.^[18][19] The repository contains the Rust source for the grok CLI/TUI and agent runtime, including the agent loop, context assembly and tool-call dispatch, file and terminal tools, the plan and diff interfaces, and the extension system for skills, plugins, hooks, MCP servers, and subagents. The public tree is periodically synchronized from SpaceXAI's internal monorepo rather than developed as a conventional community project; a SOURCE_REV file records the corresponding internal revision, and the repository says external contributions are not accepted.^[19]

First-party code in the repository is under the Apache License 2.0. Vendored and third-party code remains under its original licenses.^[19] The release is an open-source release of the client harness and UI, not of Grok model weights. Prebuilt binaries remain available for macOS, Linux, and Windows, while source builds require Rust and the repository's pinned toolchain; its README describes source builds on macOS and Linux as supported and Windows source builds as best-effort.^[19]

The open-source client also changed the model boundary. SpaceXAI says users can compile the harness, point it to their own local inference endpoint, and control it through config.toml.^[18] Current documentation supports an arbitrary custom model ID and base URL, selected with -m or /model, while Grok 4.5 became the default hosted model on July 16.^[2][26] This makes a fully local setup possible, but does not make every Grok Build session offline: browser-authenticated sessions use SpaceXAI's inference proxy, API-key sessions can call a hosted endpoint, and only a user-supplied local inference service keeps model execution local. Tool execution and local session history stay on the user's machine; for hosted inference, the documentation says prompt and selected file content are assembled locally and sent over TLS to the model service.^[28]

Privacy and security

The open-source announcement followed a privacy controversy. On July 14, Axios reported that a security researcher had observed Grok Build sending entire code repositories to a SpaceXAI-controlled Google Cloud bucket, including much more data than one test task required. The publication said the uploads appeared to stop without a client update and reported SpaceXAI's promise to delete previously uploaded user data. It also reported that customers covered by zero-data-retention agreements were said not to have trace or code data retained.^[29] The number of affected users, the full range of affected versions, how long data was stored, whether it was accessed, and independent verification of deletion remained unknown.^[29]

SpaceXAI's enterprise documentation, updated July 21, describes the current hosted path as local prompt and file assembly, TLS transport to an inference proxy, local tool execution, and local history under ~/.grok/. It says team-level zero data retention prevents prompts, code, and responses from being persisted at the inference layer.^[28] Those controls apply to the configured hosted service; a custom or local model endpoint has the data practices of that endpoint. Independent review of the now-public client improves visibility into client behavior, but it does not by itself audit server-side services or prove deletion of earlier uploads.

Marketplace, goals, and workflows

Between June and July 2026, Grok Build expanded from individual coding sessions into an extension catalog and two kinds of long-running orchestration. The plugin marketplace launched on June 11, /goal was announced on June 22, and workflows were enabled by default in CLI version 0.2.111 on July 22 before their public announcement on July 23.^{[20][22][23][24]}

Plugin marketplace

The built-in marketplace lets users browse, install, update, enable, and disable packages from the TUI or CLI. A plugin can combine skills, slash commands, agent definitions, hooks, MCP-server configuration, and language-server configuration.^[20][25] The launch catalog included packages from MongoDB, Vercel, Sentry, Chrome DevTools, Cloudflare, and Superpowers; Vercel separately confirmed that its package could be installed through /marketplace.^[20][30] Catalog sources can also be added or removed with grok plugin marketplace commands, and publishers can propose entries through pull requests to the public xai-org/plugin-marketplace catalog.^[20][21][25]

Remote catalog entries must be pinned to a full Git commit SHA, which Grok Build verifies after cloning, and catalog changes require validation and code-owner review.^[20][21] The pin protects an installation from silently moving to different code, but it is not a security audit. The marketplace repository states that third-party packages are not authored, controlled, endorsed, or verified by xAI, may execute code and access local data, and are governed by their own licenses and terms.^[21] Users therefore still need to inspect a package's contents, requested credentials, hooks, MCP servers, and upstream license before accepting the install trust prompt.

Goal mode

SpaceXAI announced /goal on June 22 for long-running autonomous work inside a session.^[22] Given an objective, the agent prepares a plan and progress checklist, executes the work, and performs its own verification, which may include code review, webpage inspection, or scripts. The user can add instructions while it runs and can inspect, pause, resume, or clear the task with /goal status, /goal pause, /goal resume, and /goal clear.^[22]

The announcement described the feature as available immediately. The changelog records goal-mode configuration in version 0.2.61 on June 22, a feature-gated welcome-screen entry in 0.2.62 on June 23, and general availability of the /goal <objective> command in version 0.2.94 on July 9.^[23] A completed checklist is the agent's record that its planned steps and verification finished; it is not independent proof that a change is correct, secure, or production-ready.

Workflows

Workflows, announced July 23 after being enabled by default in version 0.2.111, are orchestration scripts for work too broad for one conversational context.^[23][24] Grok authors phases from a natural-language request, smoke-checks the script, fans independent pieces out to agents with clean focused contexts, aggregates their results, and can add verifier agents before producing a final report. The announcement gives examples such as reviewing every feature in a large pull request, triaging 100 issues, or auditing every route for a particular flaw.^[24]

A normal workflow run receives a budget of 128 agents, with up to 1,024 for larger jobs. Progress is saved so a paused or interrupted run can resume without repeating completed work, and /workflows displays phases, agents, and per-agent token counts.^[24] Reusable workflows can be saved under a project's .grok/workflows/ directory for team use or under ~/.grok/workflows/ for one user; each saved workflow becomes a slash command with arguments. Grok Build also ships a /deep-research workflow.^[24]

Goal mode and workflows solve different problems. A goal keeps one autonomous implementation task moving against a checklist, whereas a workflow explicitly decomposes a broad job across many agent contexts and combines their outputs. In both cases, verification is part of the generated procedure rather than a formal guarantee. Large fan-outs can multiply model usage and tool actions, and SpaceXAI's workflow announcement does not publish a separate workflow price; actual access, rate limits, and cost depend on the account and model provider used for the run.

Benchmarks

The model behind Grok Build, grok-build-0.1, is a generative AI model trained specifically for agentic software engineering. xAI lists it with a 256,000-token (256K) context window and support for tool calling, prompt caching, and image inputs.^[14][15] (A few launch-day write-ups cited a much larger context figure or attributed the tool to a "Grok 4.3 beta" model; xAI's own documentation and the major API aggregators consistently list the dedicated grok-build-0.1 model with a 256K window, which is the figure used here.)

On capability, the most widely repeated number is a reported score of about 70.8% on SWE-bench Verified, a benchmark that measures whether an agent can resolve real GitHub issues.^[5] This figure appeared in trade coverage such as DevOps.com and was echoed across secondary outlets and the tool's Wikipedia entry.^[5][17] It should be read with care: at least one model-listing service noted that xAI had not formally published benchmarks for this exact variant, so the 70.8% is best described as a reported result rather than an official xAI headline number. In context, that score trailed frontier competitors. DevOps.com, for instance, contrasted Grok Build's roughly 70.8% with a markedly higher Claude Code result (reported around 87.6%), and commentators noted the gap fell short of Musk's April prediction that the tool would rival a top-tier model by late spring.^[5] Reported competitor figures varied between sources, so exact rival numbers should be treated as approximate.

Availability and pricing

Grok Build launched as an early beta and was initially delivered through xAI's consumer subscriptions rather than as a free standalone download. At launch it was gated to the SuperGrok Heavy tier, listed around $300 per month, with reporting describing a time-limited introductory rate near $99 per month for an initial six-month window. On about May 24 to 25, 2026, xAI expanded access to all SuperGrok subscribers (roughly $30 per month) and X Premium+ subscribers (roughly $40 per month), a substantial reduction in the cost of entry that moved the tool from an enthusiast-only tier toward the broader developer base.^[11]

Separately, the grok-build-0.1 model is available to developers in early access on the xAI API and through third-party gateways such as OpenRouter.^[13] Listed API pricing was approximately $1 per million input tokens and $2 per million output tokens, with cached input priced far lower (roughly $0.10 to $0.20 per million tokens), positioning grok-build-0.1 as a relatively inexpensive coding model on a per-token basis.^[13][14]

By July, the client itself could be downloaded or compiled under Apache 2.0, and SpaceXAI's product page described Grok Build as available to try free.^[3][18][19] The source license does not include free hosted inference. On July 16, SpaceXAI made Grok 4.5 the default and offered its use in Grok Build free for a limited time, without publishing an end date.^[26] The official API price for Grok 4.5 was $2 per million input tokens and $6 per million output tokens below 200,000 tokens, rising to $4 and $12 respectively for long-context requests. The older grok-build-0.1 remained listed at $1 input and $2 output below 200,000 tokens, or $2 and $4 for long context.^[27] Local models and third-party endpoints have their own hardware or provider costs.

Scroll sideways for more →

Significance

Grok Build marks xAI's formal entry into the agentic coding market, turning the 2026 contest into a three-way race among xAI, Anthropic, and OpenAI. Its arrival is notable less for raw benchmark leadership, where it trailed, than for two strategic signals. First, it shows xAI building dedicated developer products and a purpose-trained coding model (Grok 4.3 and the consumer Grok assistant aside), rather than relying on a single general model. Second, its Git-worktree-per-subagent design is a genuinely differentiated take on parallel agents: where some rivals run subagents in a shared workspace, Grok Build isolates each line of work on its own branch, a pattern that maps naturally onto how human teams use version control.

The reception was mixed but engaged. Outlets credited the parallel architecture, the plan-first workflow, and the aggressive pricing after the late-May expansion, while flagging the SWE-bench gap, the beta-stage rough edges, and the fact that competitors held large head starts in adoption and revenue. Whether Grok Build closes the capability gap that Musk himself acknowledged, and whether features like Arena Mode ship and prove useful, will determine how far it moves beyond a credible first attempt. The open-source harness, provider-independent model configuration, marketplace, goals, and large-scale workflows materially broadened that first attempt, but they also shifted evaluation toward extension trust, data handling, orchestration cost, and the reliability of agent-generated verification.

References

1. ^xAI. "Introducing Grok Build." x.ai. x.ai/...grok-build-cli
2. ^xAI Docs. "Getting Started (Grok Build overview)." docs.x.ai. docs.x.ai/...overview
3. ^xAI. "Grok Build Beta." x.ai. x.ai/cli
4. ^Engadget. "xAI introduces its coding agent called Grok Build." engadget.com/...xai-coding-agent-grok-build
5. ^DevOps.com. "xAI Enters the Coding Agent Race With Grok Build." devops.com/...he-coding-agent-race-with-grok-build
6. ^CIO Dive. "xAI joins crowded coding agent race with Grok Build." ciodive.com/...820422
7. ^eWeek. "Elon Musk's xAI Launches 'Grok Build,' a Coding Agent for Developers." eweek.com/...xai-grok-build-coding-agent
8. ^TECHi. "Grok Build turns xAI into an AI coding-agent contender." techi.com/grok-build-xai-coding-agent
9. ^Dataconomy. "xAI Launches Grok Build Coding Agent For Developers." dataconomy.com/...uild-coding-agent-for-developers
10. ^Android Headlines. "xAI Unveils Grok Build: An Agentic AI Coding Tool to Take on OpenAI, Google and Anthropic." androidheadlines.com/...ai-coding-tool-launch-beta
11. ^The Tech Outlook. "xAI launches an early beta of Grok Build; available for all SuperGrok and X Premium Plus subscribers." thetechoutlook.com/...d-x-premium-plus-subscribers
12. ^fonearena. "xAI rolls out Grok Build coding agent with ACP support and parallel subagents." fonearena.com/...-grok-build-coding-agent-features
13. ^OpenRouter. "xAI: Grok Build 0.1." openrouter.ai/...grok-build-0.1
14. ^Requesty. "grok-build-0.1 API: Pricing, Benchmarks and Specs." requesty.ai/...grok-build-0.1
15. ^KuCoin. "xAI Launches grok-build-0.1 for Agentic Coding with a 256k Context Window." kucoin.com/...ntic-coding-with-256k-context-window
16. ^Slashdot. "Elon Musk's xAI Launches 'Grok Build', Its First AI Coding Agent." developers.slashdot.org/...s-first-ai-coding-agent
17. ^Wikipedia. "Grok Build." en.wikipedia.org/...Grok_Build
18. ^xAI. "Grok Build is Now Open Source." July 15, 2026. x.ai/...grok-build-open-source
19. ^xAI. "xai-org/grok-build: SpaceXAI's coding agent harness and TUI." GitHub. github.com/...grok-build
20. ^xAI. "Grok Build Plugin Marketplace." June 11, 2026. x.ai/...grok-plugin-marketplace
21. ^xAI. "xai-org/plugin-marketplace: Official xAI plugin marketplace." GitHub. github.com/...plugin-marketplace
22. ^xAI. "Introducing /goal." June 22, 2026. x.ai/...introducing-goal
23. ^xAI. "Grok Build Changelog." x.ai/...changelog
24. ^xAI. "Workflows in Grok Build." July 23, 2026. x.ai/...workflows
25. ^xAI Docs. "Skills, Plugins & Marketplaces." docs.x.ai/...skills-plugins-marketplaces
26. ^xAI. "Introducing Grok 4.5." July 16, 2026. x.ai/...grok-4-5
27. ^xAI Docs. "Pricing." docs.x.ai/...pricing
28. ^xAI Docs. "Enterprise Deployments." docs.x.ai/...enterprise
29. ^Axios. "SpaceXAI wipes customer data after Grok uploads." July 14, 2026. axios.com/...-9168e100-7af2-11f1-bc32-bbfb768a7518
30. ^Vercel. "Vercel plugin is now available in Grok Build." June 11, 2026. vercel.com/...lugin-is-now-available-in-grok-build