CODING AGENT MIGRATION GUIDE

Grok Build 2.0 (Cursor Rebrand): Skills, AGENTS.md & MCP Migration Guide

As SpaceX and xAI consolidate Cursor's editor and CLI into the unified Grok Build 2.0 preview channel (CLI 2026.11.19), teams can keep rules and tools portable by centering on AGENTS.md, modular SKILL.md packages, and least-privilege MCP servers.

Updated 2026-11-22 · Official references: xAI · Cursor · MCP Spec.

1. What changed between Cursor and the Grok Build 2.0 unified channel

Following SpaceX's acquisition of Cursor, developer pre-releases shipped through the Cursor CLI (2026.11.19) point to a consolidated Grok Build 2.0 channel. Rather than keeping Cursor's IDE/CLI stack separate from xAI's terminal TUI coding agent (grok), Grok Build 2.0 merges both into a single authentication and orchestration surface.

For engineering teams managing shared rules, SKILL.md workflows, and MCP servers, the main operational question is portability: how to test Grok Build 2.0 without breaking developers still on legacy Cursor builds or parallel harnesses like Claude Code and Codex CLI.

  • Dual authentication in preview: unified builds support signing in with either a Grok account or a legacy Cursor account, plus headless CI/CD execution via XAI_API_KEY.
  • Multi-model routing: preview builds expose roughly three dozen selectable models across native xAI Grok models, third-party frontier models (such as Claude and Gemini), and Cursor's Composer.
  • Preview caveat: channel names, CLI flags, and subscription quotas remain in active rollout as of November 22, 2026—verify your exact CLI version and plan terms before standardizing team-wide.

2. Make rules portable: .cursor/rules to AGENTS.md and SKILL.md

Teams that previously stored all repository conventions inside editor-specific .cursor/rules/*.mdc files should promote core instructions into a root AGENTS.md file and modular SKILL.md folders. Grok Build natively reads Markdown agent instructions and auto-invokes matching skills when a task aligns with the skill description.

Keep .cursor/rules as a thin pointer (or synced build artifact) during the transition period so developers running older Cursor desktop releases and developers testing Grok Build 2.0 follow identical guardrails.

  • Place non-negotiable repo rules (test commands, branch naming, forbidden directories, secrets policy) in root AGENTS.md.
  • Split repeatable multi-step tasks (database migrations, release checks, API contract verification) into scoped SKILL.md directories with explicit trigger descriptions.
  • Use Grok Build's /skillify command to capture a verified troubleshooting session into a draft skill, then review and prune hardcoded paths before committing it to git.
Portable repo layout compatible with Grok Build 2.0, legacy Cursor, and Codex CLI
repo-root/
├── AGENTS.md                     # Single source of truth for repo rules
├── skills/
│   ├── release-preflight/
│   │   └── SKILL.md              # Scoped checklist + exact verification commands
│   └── api-schema-check/
│       └── SKILL.md
└── .cursor/
    └── rules/
        └── 00-agents-pointer.mdc # Points legacy Cursor builds to AGENTS.md

Avoid duplicating long policies in both .cursor/rules and AGENTS.md; drift between copies causes conflicting agent edits.

3. Connect and audit MCP servers in Grok Build 2.0

Grok Build includes native Model Context Protocol (MCP) support alongside hooks and plugins. You can attach remote Streamable HTTP MCP endpoints directly from the CLI or register local stdio servers for repository-scoped utilities such as git inspection, test runners, or documentation lookup.

Because unified coding agents can switch between multiple model families in one session, enforce least-privilege credentials at the MCP server boundary rather than trusting prompt instructions alone.

  • Start by registering one read-only MCP server (docs, issue tracker read, or schema inspection) and verify tool discovery before connecting write-capable servers.
  • Keep API tokens in environment variables or a secret manager rather than committing raw tokens into shared workspace MCP JSON files.
  • Pin third-party npx or uvx MCP server package versions so upstream updates cannot silently alter tool schemas mid-sprint.
Register a remote HTTP MCP server and verify headless API execution
# Add a remote HTTP MCP server in Grok Build CLI
grok mcp add --transport http docs-search https://mcp.example.com/mcp

# Headless / CI usage with xAI API key (keep key in CI secrets, never in git)
export XAI_API_KEY="<your-xai-api-key>"
grok --help

Confirm exact subcommand flags in your installed CLI version (2026.11.19+ preview) via grok mcp --help before scripting CI jobs.

4. Billing boundaries: subscription pools vs XAI_API_KEY vs third-party models

One common migration surprise is mixing subscription-bundled chat/build usage with pay-per-token API calls or third-party model billing. Grok Build access can be bundled under X / xAI subscription tiers or billed per token through the xAI API for headless automation.

When routing to non-xAI models (such as Claude or Gemini) or connecting external gateways like OpenRouter, usage may draw from separate prepaid credits or model-specific rate tables. Audit your first pilot week's billing dashboard before rolling out to a larger engineering group.

  • Separate interactive developer seats (subscription pool) from CI/CD headless runs (dedicated XAI_API_KEY with explicit spend cap).
  • Verify whether third-party frontier models inside the Grok Build 2.0 model picker use separate credits or rate limits on your plan.
  • Document fallback model settings in AGENTS.md so rate-limit fallback does not silently drop required tool-calling fidelity.

5. Pre-rollout verification checklist for teams

Run a bounded pilot on one non-production branch before switching default editor tooling. Confirm authentication, skill loading, MCP tool approval prompts, and git diff cleanliness in a real repository.

  • Verify login isolation: confirm work repositories use company-approved Grok/Cursor accounts rather than personal X accounts.
  • Test a read-only MCP call and a file-edit dry run to verify approval prompts trigger on shell execution and destructive writes.
  • Confirm /skillify output contains zero personal credentials, local usernames, or hardcoded private endpoints before sharing in git.
  • Keep legacy Cursor and alternative CLI harnesses documented until Grok Build 2.0 exits preview with finalized enterprise terms.

Frequently asked questions

Did Cursor rename to Grok Build 2.0?

Following SpaceX's acquisition of Cursor, developer pre-releases via the Cursor CLI (2026.11.19) point to a unified Grok Build 2.0 channel merging Cursor's editor/CLI stack with xAI's terminal grok agent. Because the rollout is in preview, confirm current official release notes before migrating production teams.

Can I keep using my existing Cursor account and third-party models?

Unified preview builds support sign-in with either a Grok account or a legacy Cursor account and expose an expanded model pool including Grok models, Claude, Gemini, and Composer. Confirm active plan entitlements and third-party credit rules in your account settings.

How do I connect MCP servers to Grok Build 2.0?

Use grok mcp add --transport http <name> <url> for remote HTTP MCP servers (or configure local stdio servers), starting with read-only tools and pinned versions.

How should teams migrate .cursor/rules without breaking legacy Cursor users?

Promote shared rules to root AGENTS.md and modular SKILL.md folders, and leave a lightweight pointer in .cursor/rules/ so both Grok Build 2.0 and older Cursor versions stay aligned.

Related guides