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]
Compare, price, and deploy Grok Build 0.1
These sourced facts are shared with the comparison and planning tools. Each field retains its own review date.
Grok Build 0.1
Article updated after the pricing / specification review. Inspect the sources before relying on these fields; a newer edit does not necessarily change the model facts.
- Provider API lifecycle
- ActiveSource · checked 2026-09-11 · review due
- Context tokens
- 256KSource · checked 2026-07-23
- Input / output per 1M tokens
- $1 / $2Source · checked 2026-10-25 · effective 2026-07-23
xAI announced grok-build-0.1 on its API in public beta on May 29, 2026 (its May release notes call it early access). The official model page lists no maximum output limit or committed retirement date, and alias mappings can change, so use the explicit model ID.
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 replaced grok-code-fast-1: xAI's migration guidance routes requests for grok-code-fast-1 to grok-build-0.1 after May 15, 2026.[2][8] 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 (the brand xAI used inside SpaceX, renamed SpaceXSI in October 2026) made Grok 4.5 the default model on July 16, 2026. Grok 4.7 followed on September 21, 2026, and by October the product page described Grok Build as "now powered by Grok 4.7"; current versions can also use custom model endpoints.[2][3][26][31]
xAI first teased the tool in April 2026, when Musk set a "next week" launch timeline in mid-April that slipped; the tool had not shipped by late April.[5] 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. For larger tasks it delegates work to specialized subagents that run in parallel, and subagents can be launched in their own Git worktrees, isolated checkouts of the repository, so that parallel agents cannot overwrite each other's files.[1][12][32] Worktree isolation is an option rather than the default for every subagent. A launch-day DevOps.com report said the tool would run up to eight parallel agents;[5] xAI's launch materials do not state a fixed limit.
Work flows through a plan-first model, which launch-day coverage summarized as plan, search, and build.[1][5] 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]
Launch-day coverage 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.[5] xAI's launch announcement did not mention it, and it does not appear in the Grok Build changelog through October 2026, so it should be treated as a reported capability rather than a shipped one.[1][23]
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 and Grok 4.7 had taken over that role by October 2026.[2][3][26][31] 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, xAI has not published a benchmark score for grok-build-0.1, and Requesty's model listing says no benchmarks are published for this exact variant.[14] The 70.8% SWE-bench Verified figure that circulated at launch, and was repeated by secondary outlets and Wikipedia, comes from DevOps.com, which gave it as the score of grok-code-fast-1, the 2025 model that grok-build-0.1 replaced, not as a result for the new model.[5][17]
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.
| Attribute | Detail |
|---|---|
| Product | Grok Build (agentic coding CLI) |
| Developer | xAI |
| Announced | Mid-May 2026 (announcement reported May 15, 2026)[9] |
| Wider rollout | About May 24 to 25, 2026 (all SuperGrok and X Premium+ subscribers)[11] |
| Launch model | grok-build-0.1 (replaced grok-code-fast-1)[2][8] |
| Default model | Grok 4.5 from July 16, 2026; Grok 4.7 by October 2026; custom endpoints are supported[2][3][26][31] |
| Open-source release | July 15, 2026; Rust harness and TUI under Apache 2.0[18][19] |
| Context window | 256K tokens[15] |
| SWE-bench Verified | No published score for grok-build-0.1; the widely repeated 70.8% is grok-code-fast-1's figure[5][14] |
| Parallel subagents | Yes, optionally isolated in their own Git worktrees (a launch-day report said up to 8)[1][5][32] |
| Original workflow | Plan, search, build (reviewable plan before execution) |
| Long-running modes | /goal for checklist-driven autonomous tasks; workflows for multi-phase fan-out across up to 128 agents by default and 1,024 for larger jobs[22][24] |
| Extensions | Built-in plugin marketplace launched June 11, 2026[20] |
| Interfaces | Interactive TUI, headless (-p) for CI, Agent Client Protocol (ACP)[12] |
| API pricing | About $1 / 1M input, $2 / 1M output, ~$0.10 to $0.20 / 1M cached[13] |
| Launch and May rollout subscription tiers | SuperGrok ( |
| Main rivals | Anthropic Claude Code, OpenAI Codex CLI |
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, for which xAI published no figures, 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 optional worktree isolation for parallel subagents gives a line of work its own checkout of the repository, 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 capability gap Musk had acknowledged, 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.
Continue reading
- xAIxAI is an artificial intelligence business established in 2023 under Elon Musk, who led it before the SpaceX acquisition.Mentioned 31 times above
- Claude CodeClaude Code is an agentic software development product made by Anthropic. It uses models from the Claude family to inspect software projects, propose and apply changes, run development tools, and report…Mentioned 3 times above
- OpenAI Codex CLIOpenAI Codex CLI is an open-source, terminal-based coding agent developed by OpenAI and released on April 16, 2025Mentioned twice above
- Artificial IntelligenceArtificial intelligence (AI) is the field of computer science dedicated to creating systems that perform tasks normally requiring human intelligence, such as learning, reasoning, recognizing patterns…Linked in the introduction
- Elon MuskElon Musk is the entrepreneur with the broadest operational footprint in modern artificial intelligence: he is a co-founder of OpenAI, the founder of xAI (maker of the Grok chatbot), which he led as CEO until…Linked in the introduction
- AI AgentsAn AI agent is a software system that selects and performs actions in an environment in pursuit of an objective.Linked in the introduction
More to read
Topics:AI AgentsAI Code Generation
References
- ^1 ^2 ^3 ^4 ^5xAI. "Introducing Grok Build." x.ai. x.ai/...grok-build-cli
- ^1 ^2 ^3 ^4 ^5 ^6 ^7xAI Docs. "Getting Started (Grok Build overview)." docs.x.ai. docs.x.ai/...overview
- ^1 ^2 ^3 ^4xAI. "Grok Build Beta." x.ai. x.ai/cli
- ^Engadget. "xAI introduces its coding agent called Grok Build." engadget.com/...xai-coding-agent-grok-build
- ^1 ^2 ^3 ^4 ^5 ^6 ^7DevOps.com. "xAI Enters the Coding Agent Race With Grok Build." devops.com/...he-coding-agent-race-with-grok-build
- ^CIO Dive. "xAI joins crowded coding agent race with Grok Build." ciodive.com/...820422
- ^eWeek. "Elon Musk's xAI Launches 'Grok Build,' a Coding Agent for Developers." eweek.com/...xai-grok-build-coding-agent
- ^1 ^2 ^3TECHi. "Grok Build turns xAI into an AI coding-agent contender." techi.com/grok-build-xai-coding-agent
- ^1 ^2Dataconomy. "xAI Launches Grok Build Coding Agent For Developers." dataconomy.com/...uild-coding-agent-for-developers
- ^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
- ^1 ^2 ^3The 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
- ^1 ^2 ^3fonearena. "xAI rolls out Grok Build coding agent with ACP support and parallel subagents." fonearena.com/...-grok-build-coding-agent-features
- ^1 ^2 ^3OpenRouter. "xAI: Grok Build 0.1." openrouter.ai/...grok-build-0.1
- ^1 ^2 ^3 ^4Requesty. "grok-build-0.1 API: Pricing, Benchmarks and Specs." requesty.ai/...grok-build-0.1
- ^1 ^2KuCoin. "xAI Launches grok-build-0.1 for Agentic Coding with a 256k Context Window." kucoin.com/...ntic-coding-with-256k-context-window
- ^Slashdot. "Elon Musk's xAI Launches 'Grok Build', Its First AI Coding Agent." developers.slashdot.org/...s-first-ai-coding-agent
- ^Wikipedia. "Grok Build." en.wikipedia.org/...Grok_Build
- ^1 ^2 ^3 ^4xAI. "Grok Build is Now Open Source." July 15, 2026. x.ai/...grok-build-open-source
- ^1 ^2 ^3 ^4 ^5 ^6xAI. "xai-org/grok-build: SpaceXAI's coding agent harness and TUI." GitHub. github.com/...grok-build
- ^1 ^2 ^3 ^4 ^5 ^6xAI. "Grok Build Plugin Marketplace." June 11, 2026. x.ai/...grok-plugin-marketplace
- ^1 ^2 ^3xAI. "xai-org/plugin-marketplace: Official xAI plugin marketplace." GitHub. github.com/...plugin-marketplace
- ^1 ^2 ^3 ^4xAI. "Introducing /goal." June 22, 2026. x.ai/...introducing-goal
- ^1 ^2 ^3 ^4xAI. "Grok Build Changelog." x.ai/...changelog
- ^1 ^2 ^3 ^4 ^5 ^6xAI. "Workflows in Grok Build." July 23, 2026. x.ai/...workflows
- ^1 ^2xAI Docs. "Skills, Plugins & Marketplaces." docs.x.ai/...skills-plugins-marketplaces
- ^1 ^2 ^3 ^4xAI. "Introducing Grok 4.5." July 16, 2026. x.ai/...grok-4-5
- ^xAI Docs. "Pricing." docs.x.ai/...pricing
- ^1 ^2xAI Docs. "Enterprise Deployments." docs.x.ai/...enterprise
- ^1 ^2Axios. "SpaceXAI wipes customer data after Grok uploads." July 14, 2026. axios.com/...-9168e100-7af2-11f1-bc32-bbfb768a7518
- ^Vercel. "Vercel plugin is now available in Grok Build." June 11, 2026. vercel.com/...lugin-is-now-available-in-grok-build
- ^1 ^2 ^3xAI. "Introducing Grok 4.7." September 21, 2026. x.ai/...grok-4-7
- ^1 ^2xAI Docs. "Worktrees." docs.x.ai/...worktrees
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
8 revisions · v8 · 3,224 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check of the full article against its cited sources (110 claims, 33 sources fetched); 6 material + 10 minor fixed
Cite this page: AI Wiki. "Grok Build." aiwiki.ai, updated 1 Nov 2026, fact-checked 5 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/grok_build



