Grok Build, Grok Skills + Connectors: xAI Dev Stack 2026

For about eighteen months, xAI's developer story was "Grok is a chat model, and Grok 4 has a generous context window." That stopped being true in May 2026. Across roughly two weeks, xAI shipped a coding agent, a skills system, and a connectors layer — the three pieces every serious AI-developer platform now has. This is xAI making a deliberate move from model to stack.

If you're evaluating where Grok fits next to Claude Code, Cursor, or GitHub Copilot Workspaces, this is the piece for you. We'll go product by product, then look at how they snap together, and finish with a sober comparison against the incumbents.

What is Grok Build?

Grok Build launched in early beta on May 14, 2026 as xAI's first agentic coding CLI — a terminal-native coding agent in the same family as Claude Code and OpenAI's Codex CLI. It is powered by grok-code-fast-1 (xAI's purpose-trained coding model from August 2025) and the newer grok-build-0.1 API model, published on May 20, 2026, with a 256,000-token context window and image-plus-text input.

How does Grok Build actually work?

The headline mechanic is parallel agents. A single prompt can spawn up to eight sub-agents, each working on its own branch of the codebase. The UI exposes two underlying models — Grok Code 1 Fast and Grok 4 Fast — with up to four agents per model. You stop waiting for one agent to finish; you fan out, then pick the winning branch.

Each agent walks a three-stage workflow:

1. Plan — produces an explicit, step-by-step execution plan before touching files. You can approve the whole plan, comment on individual steps, or rewrite parts before a single line of code is written.
2. Search — reads the relevant files, navigates the repo, builds working memory.
3. Build — executes the approved plan, runs commands, writes code.

The plan-mode gate is the part long-time Claude Code and Aider users will recognise immediately: it's the same "no edits without a plan you accepted" pattern, just enforced in the UI rather than as a convention.

What is Arena Mode?

Arena Mode is the layer that makes 8-parallel-agents actually usable. Instead of dumping eight raw outputs in front of you and asking which one is best, Arena scores the competing solutions on an automated evaluation pass (tests pass? diff size? plan adherence?) and ranks them before they hit your review queue. Arena was confirmed in code traces earlier in 2026 and is rolling out across the beta — it's not yet on by default in every install.

What does Grok Build cost?

Two routes:

• Subscription: bundled into SuperGrok and X Premium Plus plans. Deeper agentic features sit on SuperGrok Heavy ($300/mo), and xAI launched a separate SuperHeavy tier ($299/mo, with a $99/mo introductory promo for the first six months) that includes the full Grok Build beta.
• API: $1 per million input tokens, $2 per million output tokens, with $0.20/M cached input. Pricing is for the grok-build-0.1 model (256K context, text + image input).

For context, the coding model powering Grok Build (grok-code-fast-1) scored 70.8% on SWE-Bench Verified on xAI's internal harness, which puts it roughly 15–18 points below Claude Opus 4.7 and GPT-5.5 on the most-cited agentic-coding benchmark — respectable for a v1 release, behind the frontier. We covered the head-to-head in detail in Grok Build vs Claude Code vs Codex CLI, and the install path in How to install Grok Build.

What are Grok Skills?

Grok Skills, launched on May 18, 2026, is xAI's answer to the "persistent custom expertise" pattern that Anthropic introduced with Claude Skills and OpenAI with custom GPTs.

The model is simple: you teach Grok how to do a task once — the format, the tone, the workflow steps — and Grok carries that across every future conversation. No more re-pasting the same system prompt. A Skill is a small bundle (name, description, instructions; optional reference files) that you upload, then invoke with a slash command or by intent.

Are Grok Skills compatible with Anthropic Skills?

This is the interesting part for developers already invested in the Anthropic Skills ecosystem. xAI's official documentation states that "Grok is fully compatible with Claude Code with zero configuration needed." The Grok Build docs spell out that Grok automatically reads Claude Code marketplaces, plugins, skills, MCPs, agents, hooks, and instruction files (CLAUDE.md, Claude.md, CLAUDE.local.md, and .claude/rules/) alongside Grok's own .grok/ directory.

The Grok Skills upload dialog accepts .zip, .skill, and .md files — the same shapes Anthropic uses for Claude Skills. Most Skills written for Claude Code — a SKILL.md with YAML frontmatter and a markdown body — load into Grok with no edits, and the community is already swapping them.

If you're writing internal Skills today, this matters: design them as if they need to run in two stacks. Use generic terminology, avoid stack-specific tool calls in the front-matter, and keep the instructions declarative rather than imperative. We have a full primer on writing portable agent context in our AGENTS.md guide.

What tools do Grok Skills ship with?

Skills come with a built-in toolset: full Word-document generation and editing (with preserved headings, tables, and styles), PowerPoint-style slide-deck creation with speaker notes, Excel spreadsheets with formulas, charts, and conditional formatting, PDF operations, and a workflow-automation runner. Skills are available on Grok 4.3 across web, iOS, and Android — not just inside Grok Build. So a "PR-review checklist" Skill works whether you're in the terminal coding agent or pasting a diff into the web app.

What are Connectors?

Connectors landed in two waves. The initial launch was May 6, 2026, with GitHub, Notion, Linear, Google Workspace, Microsoft 365, and Bring-Your-Own-MCP. A second wave on May 22, 2026 added Vercel, Canva, Gamma, and S&P Global. Connectors are first-party integrations that give Grok read/write access to the apps developers and teams actually live in.

What connectors does Grok support today?

The combined lineup as of late May 2026 covers most of the obvious surface area:

• GitHub — read and modify repository content; powers Grok Build's repo operations.
• Notion — read databases, create new pages, edit blocks via natural-language commands.
• Linear — issue queries and updates.
• Google Workspace — Gmail, Drive, Docs, Sheets, Calendar through a single integration point.
• Microsoft 365 — SharePoint, Outlook, OneDrive, Teams.
• Salesforce — CRM record queries and updates.
• Vercel — build and deploy sites without leaving Grok (added May 22).
• Canva — create designs from prompts (added May 22).
• Gamma — assemble presentations (added May 22).
• S&P Global — live market data for finance workflows (added May 22).
• X — native real-time post and account access.
• Grokipedia — xAI's in-house knowledge base.

Can I bring my own MCP server?

Yes — and this is the move that signals xAI is playing the long game. Connectors supports Bring Your Own MCP: point Grok at any Model Context Protocol server you've already configured for Claude or Cursor (databases, internal APIs, observability tools) and it shows up alongside the first-party connectors. The protocol is the same one Anthropic shipped in late 2024, now multi-vendor in practice. The one technical caveat: the MCP server must be publicly internet-accessible — local-only deployments need a tunnel.

Connectors are live on web, iOS, and Android. Most require a SuperGrok tier; advanced ones (deep agentic write access, scheduled runs) sit behind SuperGrok Heavy.

How do Grok Build, Skills, and Connectors work together?

Looked at one-by-one, each piece is a respectable v1. The interesting part is how they compose:

• A Skill defines the workflow. "When asked to ship a feature: read the Linear ticket, check out the right branch, run plan mode, draft the PR description in our house style, open the PR."
• Connectors provide the surface. The Linear connector reads the ticket. The GitHub connector handles the branch and PR. The Notion connector posts the changelog.
• Grok Build does the work. The coding agent does the planning, searching, and building — with up to eight parallel attempts and Arena ranking.

That's the same end-to-end loop Claude Code + Anthropic Skills + Anthropic Connectors offers today, and the same one Cursor + Cursor Agents + integrations is reaching for. xAI's edge, where it exists, is parallel-agent throughput, the SuperGrok subscription bundle, and tight native access to X and Grokipedia for teams that already live there. Where it lags is reasoning depth on complex multi-file changes and the maturity of the Skills marketplace.

How does the xAI stack compare to Claude Code, Cursor, and Copilot Workspaces?

Honest read on each axis:

Reasoning depth on complex tasks

Claude Code (Opus 4.7) leads. On multi-file refactors and large-codebase bug fixes, Opus's reasoning and 200K-context working memory remain the strongest in the field. Grok Build is faster on isolated, well-scoped fixes but trails on the gnarly stuff — the SWE-Bench gap (70.8% vs ~87–88%) shows up in the field.

Editor experience

Cursor wins. A VS Code fork with AI baked into the autocomplete and Composer is still the fastest day-to-day surface for most developers. Grok Build is terminal-only today, like Claude Code. We have a deep dive in our Cursor IDE guide.

Parallelism

Grok Build leads, narrowly. Eight parallel agents with automated ranking is genuinely novel; Claude Code does sub-agents but not yet with first-class arena scoring. Cursor and Copilot Workspaces don't have a direct equivalent.

Skills and marketplace

Anthropic leads on ecosystem; xAI rides on top. Claude Skills launched earlier and has more community packs. Grok's decision to read the Anthropic format directly — including CLAUDE.md, .claude/rules/, plugins, and skills — is a clever fast-follow that imports the ecosystem rather than competing with it.

Connectors and integrations

Roughly at parity. Anthropic, OpenAI, and xAI all support MCP. xAI's first-party list (GitHub, Notion, Linear, Workspace, Microsoft 365, Salesforce, Vercel, Canva, Gamma, S&P Global, plus X and Grokipedia) is competitive with Anthropic's and a bit ahead of OpenAI's for design and presentation workflows.

Pricing

If you already pay for X Premium Plus or SuperGrok, the bundle math is hard to beat — you get the coding agent, Skills, and Connectors at the subscription you're already paying. API pricing ($1/$2 per million tokens, $0.20 cached input) undercuts Claude Sonnet ($3/$15) and Opus 4.7 ($15/$75) substantially. Whether the savings outweigh the SWE-Bench gap depends on your workload.

Which should you use?

Honest take: most teams should still default to Claude Code or Cursor for the primary surface today, with Grok Build as a fast-and-cheap secondary for high-volume, well-scoped work. If you're already inside the X / SuperGrok ecosystem, Grok Build + Skills + Connectors is good enough to do real work on. If you're shopping fresh, the Claude Code stack is the safer bet for the next 6–9 months while xAI catches up on the gnarly multi-file reasoning gap.

For more on the full landscape, see our complete guide to AI coding agents. And for the broader May 2026 picture (Grok 4.3, the other models, Anthropic news), see our May 2026 AI model releases roundup.

What does this mean for engineering teams?

Three concrete takeaways:

1. The "chat model + coding agent + skills + connectors" pattern is now table stakes. Every serious AI-developer platform — Anthropic, OpenAI, xAI, GitHub — ships the same four pieces. Pick on price, ecosystem fit, and reasoning quality, not on whether the pieces exist.
2. Skills are the new portable surface. If you write internal Skills today, write them in the Anthropic format. Grok reads them natively. Cursor takes a subset. OpenAI's tooling will land similarly soon. The cost of building stack-specific custom agents has dropped.
3. MCP won. Bring-your-own-MCP is now in every major dev stack. If you have internal data sources, build one MCP server, plug into all four stacks.

FAQ

Is Grok Build available to all xAI users?

No — it's in early beta for SuperGrok Heavy and SuperHeavy subscribers, with broader rollout to SuperGrok and X Premium Plus underway. The underlying grok-build-0.1 model is also available via the xAI API at $1 / $2 per million input/output tokens for any developer with API access.

What benchmark score does Grok Build get?

The coding model that powers it (grok-code-fast-1) scored 70.8% on SWE-Bench Verified on xAI's internal harness. That puts it behind Claude Opus 4.7 (around 87–88%) and GPT-5.5 (similar range), but ahead of most older open-weight coding models. It's a respectable v1.

Are Grok Skills compatible with Anthropic Skills?

Yes — xAI's documentation explicitly states "Grok is fully compatible with Claude Code with zero configuration needed." Grok automatically reads Claude Code marketplaces, plugins, skills, MCPs, agents, hooks, and instruction files (CLAUDE.md, Claude.md, CLAUDE.local.md, .claude/rules/). The Grok Skills uploader accepts .zip, .skill, and .md files matching Anthropic's SKILL.md convention. Test before relying on it for production, but the format compatibility is real and intentional.

How many parallel agents can Grok Build run?

Up to 8 sub-agents in parallel — the UI exposes two underlying models (Grok Code 1 Fast and Grok 4 Fast) with up to 4 agents per model. Arena Mode ranks the outputs automatically before they reach your review queue.

What connectors are available?

The May 6 initial wave included GitHub, Notion, Linear, Google Workspace, Microsoft 365 (SharePoint, Outlook, OneDrive, Teams), Salesforce, and Bring Your Own MCP. The May 22 expansion added Vercel, Canva, Gamma, and S&P Global. X and Grokipedia round out the first-party set.

Does Grok Build keep code on my machine?

Yes — Grok Build is built local-first. Code runs on your machine, and nothing in your codebase is transmitted to xAI's servers during a session (only per-prompt tokens for inference). It's also air-gap compatible once the initial setup completes. Treat the trust boundary the same as Claude Code.

How does plan mode compare to Claude Code's planning behaviour?

Conceptually identical — the agent produces an execution plan before editing files, and waits for you to approve or modify. Grok Build makes this a first-class UI step you can't accidentally skip; Claude Code makes it a strong convention but is more flexible about how you opt in or out.