Definition

Model Context Protocol (MCP) is an open-source protocol developed by Anthropic that enables AI models to connect to external tools, data sources, and applications. It serves as a standard for AI integration, similar to how USB provides a standard for hardware connections.

Overview

MCP provides a standardized way for AI systems to:

  • Access external data sources
  • Call APIs and services
  • Perform actions in external applications
  • Maintain context across interactions

2026 Adoption

Spec 2026-07-28 (July 2026)

Warning

~250M weekly SDK downloads and AAIF ~240 members are maintainer/foundation-cited via VentureBeat — attribute.

Anthropic Claude Expansion

  • Claude Connectors expanded to 200+ applications via MCP
  • Consumer integrations: Uber, Spotify, Uber Eats, Instacart, TurboTax, Resy, Booking.com
  • OAuth 2.0 authentication standard
  • Privacy: No persistent data storage, no training on Enterprise/Team data

Sentry MCP Integration (Security Concern)

  • sentry MCP server returns error events to ai-coding-tools as trusted structured output
  • agentjacking exploit: injected error events indistinguishable from legitimate Sentry guidance
  • Implicit trust model — agents cannot verify data provenance

Security Risks (June 2026)

RiskDescription
Implicit trustAgents treat MCP responses as authoritative system data
Event injectionExternally-influenced data (Sentry DSN, support tickets) becomes agent input
Prompt injection via MCPMalicious servers craft prompts requesting LLM completions
Session hijackingStateful MCP connections vulnerable to replay/resume attacks
Covert tool invocationHidden tool calls without user awareness

MCP sampling and tool responses create indirect prompt-injection vectors. Defense requires execution-layer controls, not prompt hardening alone.

Industry Impact

MCP is emerging as a potential standard for AI connectors:

  • Enables “plug and play” AI integrations
  • Reduces custom integration development effort
  • Expands attack surface as more tools connect via MCP (agentjacking demonstrated at scale)

Key Points

Sources