Model Context Protocol (MCP)

An open standard developed by Anthropic that defines a standardised way for AI agents to connect to external tools, data sources, and services. The analogy used in ^[src-003]: MCP is a universal plug for AI agents — instead of writing custom integrations for every tool, an agent speaks one protocol and any MCP-compatible tool just works. Rapidly becoming the industry standard for agentic AI tool connectivity in 2026 ^[src-003].

Key points

Open standard, not a product ^[src-003]. Anthropic published the specification; anyone can implement an MCP server (exposing tools) or client (consuming them). This decoupling is what makes it a standard rather than a proprietary integration.

Solves the N×M integration problem. Without MCP, every AI agent needed custom code to talk to every tool → N agents × M tools = a quadratic integration burden. With MCP, each agent speaks one protocol and each tool exposes one interface → N + M.

Widely adopted in 2026: Claude Code ships with MCP support natively. Community MCP servers exist for WordPress, Gmail, Google Calendar, Slack, Linear, GitHub, Linear, and many more. In Robin's own setup, WordPress content management and site operations are handled entirely through MCP servers.

Used by the Agent Orchestration layer, not the model itself. MCP defines how tools get registered, invoked, and how results flow back — the reasoning about which tool to call still happens inside the LLM's React Loop.
Statsig's AI experimentation article frames MCP as a way for a product's novel context to be integrated with AI, which increases the need to test and optimize models, prompts, datasets, and agent workflows in context ^[src-032].
Google's A2UI v0.9 post positions MCP as one possible transport for generative UI: A2UI can run over MCP, WebSockets, REST, AG-UI, A2A, or other transport interfaces ^[src-038].
This clarifies MCP's role in the emerging agent UI stack: MCP can carry tool/app integration, while A2ui standardizes the UI intent that the client renders ^[src-038].
Google Labs' Stitch article adds a design-tool example: the Stitch MCP server and SDK let other tools and skills access Stitch capabilities, making design canvases part of the agent toolchain ^[src-040].
Marco Mornati's MCP-vs-CLI article distinguishes MCP as a protocol from the costly native injection pattern: the problem is not structured tool calling itself, but paying the Tool Schema Tax for idle tools on every prompt ^[src-041].
The article identifies Gateway MCP Pattern as a way to retain MCP-style structured outputs while loading service-specific schemas lazily ^[src-041].
Google Cloud's enterprise-agent demo treats MCP servers as observable topology nodes: the Agent Registry can show traffic between agents, MCP tools, MCP servers, and other agents, including metrics such as latency and errors ^[src-043].
This adds an operations requirement to MCP adoption: once MCP tools are production dependencies, they need governance observability, egress checks, and trace correlation ^[src-043].
Google Cloud Next '26 pushes MCP into cloud operations: Google says it uses MCP to expose every Google Cloud service as an agent-orchestratable tool, and Agent Gateway understands MCP alongside A2A and other agentic protocols ^[src-044].
Google also says it will expose Google Cloud services as MCP so agents can interact with Google Cloud from any agent environment ^[src-044].
Boris Cherny says MCP is Anthropic's simplest answer for giving Co-work and Claude surfaces access to knowledge-work tools such as Salesforce, Google Docs, and Google Calendar; for systems without MCP, computer use acts as a slower catchall ^[src-054].
Rory Richardson adds an architecture warning: if agents become the new microservices, MCP servers may become a gateway layer, but many are not yet as hardened or fortified as mature API gateways ^[src-057].
That warning pushes MCP from developer convenience toward production infrastructure: agent-facing tool servers need structure, hardening, permissions, and verification before enterprises rely on them as core boundaries ^[src-057].
The AI Engineer corpus captures MCP's rapid move from novelty to infrastructure: talks cover full-spec MCP capabilities, protected MCP servers, remote MCPs, MCP observability, MCP registries, cross-app access, A2A plus MCP workflows, and critiques of MCP overuse ^[src-077].
The practical takeaway is that MCP is not just a connector format. Once exposed to agents, MCP servers become security, identity, observability, schema, and product-boundary decisions ^[src-077].
Fmind's MCP videos add a didactic comparison point: MCP standardizes how agents connect to tools and context providers, while Agent To Agent Protocol standardizes coordination between agents ^[src-078].
The course framing makes MCP easier to place inside the engineering stack: it is an integration layer that still needs the MLOps-style disciplines of configuration, testing, security, documentation, release management, and observability ^[src-078].
Roberts's Hermes Agent OS demo treats Zapier-style MCP connectors as a fast way to add Gmail and Calendar access to a personal agent, but with least-access permissions such as drafting rather than sending email by default ^[src-079].
^[src-088] adds two practical frontiers. Supabase's talk separates skills from MCP: skills encode product-specific judgement and context, while MCP exposes live tools and data; strong agents often need both rather than treating them as substitutes.
RL Nabors's WebMCP talk moves the pattern into the browser: web pages can expose agent-callable affordances, while MCP apps and HTTP-hosted MCP servers turn web properties into richer agent surfaces ^[src-088].
^[src-104] adds a voice-agent example: Obsidian's local MCP/API plugin lets a GPT Realtime 2 assistant create, search, read, and write notes by voice, while apps without MCP need a separate desktop-control path.

^[src-109] adds a commerce-specific MCP use case: Shopify's Catalog, Cart, and Checkout MCPs expose discovery-to-purchase actions for agentic shopping experiences via UCP.
^[src-115] adds a UI transport angle: MCP servers can return A2UI payloads such as application/a2ui+json, and MCP Apps can be embedded into A2UI surfaces when custom sandboxed UI is needed.
^[src-116] clarifies the adjacent discovery layer: Agentic Resource Discovery helps agents find and verify capabilities across registries, while MCP remains one of the protocols those capabilities may use after discovery.
Typical MCP server responsibilities:

– Declare what tools it provides (name, description, parameter schema)

– Declare what resources it exposes (readable content, e.g. file trees, database rows)

– Handle tool invocations and return structured results

– Handle resource reads

Why it matters for Robin's setup

Every skill in the .claude/skills/ folder that talks to an external system — WordPress, Gmail, Calendar — does so through an MCP server. This is why adding a new integration to Robin's assistant is cheap: if an MCP server exists, you just connect it. If it doesn't, you write one small server and every future agent can use it.

Related entities

Shopify
Claude Code — the primary runtime that consumes MCP servers
Paperclip — can use MCP tools per agent

Related concepts

Updates from bulk ingest

From src-005 (cluster 2)

n8n (v1.21.2+) exposes an instance-level MCP server with three tools: search workflows, get workflow details, execute workflow
Claude has a native n8n MCP connector; ChatGPT requires developer mode + OAuth; Lovable has a native integration
Community n8n MCP server is a separate project exposing node catalogue and templates (not live instance) – used by Claude Code for workflow building
MCP clients only see n8n workflows explicitly toggled 'available in MCP' and published – safety gate by design
Firecrawl, GitHub, and Anthropic all expose MCP servers that Claude Code can install via natural language
Claude Code installs MCP servers by writing an mcp.json file in the project with server URL and credentials – credentials are stored locally, not in GitHub repos

From src-007 (cluster 4)

Canonical real-world MCP use case: Vapi voice agent discovers and calls n8n sub-workflows through an n8n MCP server trigger, bearer-authenticated via SSE. Illustrates MCP's value for consolidating tool registries and avoiding duplicated schemas (y-cq_Qo4zVo)

2026-06-22 update

Google Cloud Tech adds two applied MCP references: building a custom MCP server for an agent ^[src-128] and using MCP to let agents access Looker's governed semantic layer ^[src-139].

Source references

^[src-109] Shopify – "Shopify Spring '26 agentic commerce release" (2026-06-17)

^[src-003] Robin Cartier — "What is Agentic AI? A Complete Guide" (2026-03-10)
^[src-005] Nate Herk cluster (see summaries/src-005-*.md)
^[src-007] Nate Herk cluster (see summaries/src-007-*.md)
^[src-032] Skye Scofield and Sid Kumar — "Experimentation and AI: 4 trends we’re seeing" (2025-06-13)
^[src-038] Google A2UI Team — "A2UI v0.9: The New Standard for Portable, Framework-Agnostic Generative UI" (2026-04-17)
^[src-040] Rustin Banks — "Design UI using AI with Stitch from Google Labs" (2026-03-18)
^[src-041] Marco Mornati — "The Future of Agentic Tooling: MCP Servers vs. CLI A Data-Driven Comparison" (2026-04-27)
^[src-043] Google Cloud Events — "Operationalize AI: A blueprint for managing enterprise agents at scale" (2026-04-24)
^[src-044] Thomas Kurian — "Welcome to Google Cloud Next '26" (2026-04-22)
^[src-054] Sequoia Capital — "Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next" (2026-05-04)
^[src-057] Amazon Web Services — "The Future of Agentic AI with Rory Richardson | AWS Humans In The Loop Podcast" (2026-05-01)
^[src-077] AI Engineer channel transcript cluster (678 saved transcripts, 2023-10-20 to 2026-05-15)
^[src-078] Mederic Hurier (Fmind) channel transcript cluster (62 saved transcripts, 2024-11-26 to 2026-05-14)
^[src-079] Jack Roberts — "Hermes Agent just got 10X Better (Agentic OS)" (2026-05-15)
^[src-088] AI Engineer late-May 2026 channel update (48 transcripts, 2026-05-15 to 2026-05-31)
^[src-104] Pat Simmons – "GPT Realtime 2 Can Now Run Your Entire Computer (Just Your Voice)" (2026-06-17)
^[src-115] Google Developers Blog – "A2UI + MCP Apps" (2026-06-17)
^[src-116] Google Developers Blog – "Announcing the Agentic Resource Discovery specification" (2026-06-17)
^[src-128] Google Cloud Tech – "How to build an AI Agent and MCP Server" (2026-06-18)
^[src-139] Google Cloud Tech – "Governed agents: Looker + MCP explained" (2026-06-19)

2026-06-27 update

Google Cloud Tech's MCP explainer reinforces that MCP is becoming baseline agent literacy; the watch focus should shift to which MCP tools are safe, observable, and worth exposing ^[src-166]. The transcript-backed upgrade confirms the explainer should be treated as evidence, not just a metadata bookmark.
AI Engineer's June 2026 World Fair slice adds MCP product-surface evidence: WebMCP, MCP apps, double iframes, browser-agent interfaces, Chrome DevTools patterns, and generative UI all move MCP from connector plumbing toward agent-facing interaction design ^[src-185].
Google Cloud Tech's ADK tutorial shows a concrete split between MCP for live web access and custom stateful tools for app-specific operations, with safety callbacks around model input, model output, and tool arguments ^[src-189].
The World's Fair playlist update adds a dedicated MCP Apps talk that reinforces MCP's move toward application primitives, discovery, and user-facing software surfaces rather than only backend connector plumbing ^[src-191].

2026-06-29 MCP adoption discipline and perimeter controls

Daniel Oh's Agentic AI Foundation talk is a useful corrective to MCP hype: if one agent calls one known tool, direct function calling may be cheaper and simpler than an MCP server ^[src-172].
Google Cloud's VPC Service Controls update shows MCP becoming policy surface area: perimeter rules can use attributes such as mcp.toolName, mcp.method, and mcp.tool.isReadOnly ^[src-174].
The combined lesson is that MCP is both an integration protocol and an operational/security boundary; teams should adopt it where discovery, multi-client access, governance, or perimeter policy justify the extra moving parts ^[src-172] ^[src-174].
^[src-191] also connects MCP Apps to the wider agentic-UX theme: MCP surfaces need receipts, discovery, rendering, and interaction design when agents and humans share the same software workflow.

Robin Cartier perspective

This page is part of Robin Cartier's working AI knowledge graph: a practical research layer for production AI, recommendation systems, experimentation, GEO, and agentic web readiness.

The useful next step is to connect this concept back to applied product leadership and operating models.

Recommended next

Keep reading from this thread

From 491 indexed pages and articles.