Skip to main content
reopt Handbook
reopt Handbook
Vercel Enterprise AI Platform

Platform Foundation

Platform ArchitectureAI SDK RuntimeAI Gateway Control Plane

Execution Runtime

Workflow Durable ExecutionSandbox Tool RuntimeMCP Enterprise DataQueues and Async Jobs

Prompt and Quality

Prompt Engineering and GuardrailsRAG and Retrieval-Augmented GenerationObservability and Evaluation

Operations

Security GovernanceCost and ReliabilityDeployment and AI CI/CD

Scenarios

Customer Support Agent ArchitectureInternal Research Agent ArchitectureApproval Backoffice AutomationCoding Orchestration

Orchestration Patterns

Graph-Centric OrchestrationPractical Orchestration PlaybookMulti-Agent Architecture

Appendix

Migration GuideReferencesVerificationUpdates
Handbook›Vercel Enterprise AI Platform›MCP Enterprise Data
한국어KOEnglishEN

MCP Enterprise Data

Connect enterprise systems to AI agents through MCP tools, resources, and prompts.

Written and reviewed by Eric Han · Co-founder & CTO, reopt

Key takeaways

  • MCP is a useful enterprise data boundary when it exposes narrow tools and resources instead of broad database or API access.
  • Design resources as read-only permission-filtered context, tools as narrow schema-bound actions, and prompts as reusable task patterns.
  • Enforce source-system permissions inside MCP servers and return structured results rather than raw dumps.
  • Separate read tools from mutation tools, require approval for side effects, and version tool schemas with deprecation notes.
  • Test the boundary by asking whether one MCP server could safely serve two teams with different permissions.

AI SDK 7 and development MCP

Use createMCPClient from @ai-sdk/mcp, load an explicit tool allowlist, and pass the resulting tools to the generation call. Close a request-scoped client after completion; keep streaming clients alive until the stream ends. Server annotations are untrusted hints, so enforce tenant credentials, permissions, and approval policy separately. See AI SDK MCP.

Next.js 16+ exposes a development MCP endpoint at /_next/mcp; next-devtools-mcp discovers it and forwards runtime tool calls. Check the installed server's tool inventory. Keep this development diagnostic connection separate from enterprise data services. See Next.js MCP.

MCP is a useful boundary for enterprise data access when it exposes narrow tools and resources instead of giving agents broad database or API access.

MCP Resource Design

CapabilityRecommended posture
ResourcesRead-only, permission-filtered context
ToolsNarrow actions with explicit schemas
PromptsReusable task patterns and constraints
AuthUser-aware or service-scoped credentials
AuditLog every tool call with actor and input

Enterprise Rules

  • Enforce source system permissions inside MCP servers.
  • Return structured results rather than raw dumps.
  • Separate read tools from mutation tools.
  • Require approval for side effects.
  • Version tool schemas and publish deprecation notes.

Review Question

Could the same MCP server safely serve two teams with different permissions? If not, the access boundary is probably too broad.

Related docs

MCP Integration

Advanced Codex Usage · Connect Codex to MCP servers with explicit tool, data, and trust boundaries.

MCP Integration

Claude Code Complete Guide · Connect Claude Code to tools and resources through Model Context Protocol servers.

Cmd. /mcp

Codex Command Master · List configured MCP tools and inspect server diagnostics.

From Docs to Skills, Plugins, and MCP

Agentic Documentation · Split rules, workflows, distribution units, and external context into the right agent surface.

Templates

Agentic Documentation · Templates for AGENTS.md, CLAUDE.md, Skills, Plugins, MCP, llms.txt, runbooks, and KB docs.

Sandbox Tool Runtime

Isolate code, file, browser, and shell execution for AI agents.

Queues and Async Jobs

Use queues for fan-out, batch processing, retries, and rate-controlled AI work.

On this page

AI SDK 7 and development MCPMCP Resource DesignEnterprise RulesReview Question