The practical companion to MCP Will Change Which Businesses AI Recommends. That article covers the why and whether. This guide covers the how and when.
Step one: understand which role applies to you
Most MCP confusion comes from conflating three distinct roles. The role that matters to a business owner is almost always the third one.
MCP Host — the AI application the user interacts with that can connect to MCP servers. Examples: Claude Desktop, ChatGPT desktop app (Developer Mode, macOS), Cursor, VS Code with Copilot, Microsoft Copilot Studio. A business owner does not build an MCP host — they choose one.
MCP Client — the protocol layer inside a host that manages connections to servers. Invisible plumbing. Not relevant to business owners.
MCP Server — the component a business builds or configures to expose its own data — products, pricing, availability, credentials — to any MCP-compatible AI system. This is the role that determines whether your business is queryable by AI agents.
The role that determines your business’s MCP eligibility is the server role — whether you have something that exposes your data to AI systems. Everything else is infrastructure you choose or configure to connect to it.
Which AI systems actually support MCP — the honest landscape
In plain terms: MCP is mature in developer tools, emerging in enterprise platforms, and not yet mainstream in consumer AI. MCP support is not binary — it ranges from full production-ready implementation to actively stepping back from the protocol. The detail matters before making any investment decision.
Claude Desktop (Anthropic) — ✅ Full support. The most mature MCP host and reference implementation. Where MCP actually works in production today. Requires desktop app install.
ChatGPT desktop (OpenAI) — ✅/⚠️ MCP via Developer Mode since March 2025. macOS desktop app only — remote MCP for the web interface still in development. OpenAI adopted MCP formally in March 2025 across Agents SDK and Responses API.
Gemini (Google) — 🔄 Building. Google DeepMind confirmed MCP support in April 2025. Google Cloud MCP servers available. Enterprise-focused — not yet mainstream for consumer use cases.
Copilot Studio (Microsoft) — 🔄 Building. Integrated into Azure OpenAI and Microsoft Semantic Kernel. Enterprise-grade but not configuration-level for most businesses yet.
Perplexity — ⚠️ Stepping back. In March 2026, Perplexity’s CTO announced moving away from MCP toward a traditional API, citing token consumption (40–50% of context windows) as the primary reason. Still runs an MCP server, but it is no longer their flagship approach.
Cursor / VS Code / Windsurf / Zed — ✅ Full support. Where MCP sees highest day-to-day use in 2026. Developer coding tools — not relevant for business-facing AI recommendation use cases. Important context: MCP’s most active current users are developers, not consumers.
Claude.ai web — 🔄 Partial. Has connectors (Google Drive, Gmail, etc.) but these are Anthropic-managed integrations, not open MCP. Full open MCP on the web interface is not yet available.
The uncomfortable truth: MCP’s most mature, production-ready implementations in March 2026 are in developer coding tools — not in the consumer AI systems most businesses picture when they imagine “AI recommending them.” ChatGPT’s MCP support is real but limited to a desktop app in Developer Mode. The consumer-facing MCP story is still being built.
MCP servers that already exist — what you can connect today
Shopify — native MCP endpoint enabled by default since Summer 2025 (expanded Winter ’26 Edition). Product catalogue, pricing, inventory, cart management, checkout. Every Shopify store already has a live MCP endpoint. This is the most commercially significant MCP implementation available to SMBs today.
WooCommerce (WordPress) — a WordPress plugin provides read-only access to products, categories, reviews, and WordPress content. Install, configure, test with Claude Desktop. Accessible today without developer resource.
HubSpot — official MCP server available. Read-only CRM data: contacts, companies, deals, tickets. Currently developer-facing — not yet configuration-level for non-technical users.
Salesforce — third-party server available with create/update/delete capability. Salesforce’s own hosted server still in beta as of early 2026.
Google Workspace — Drive, Calendar, Gmail. Google Cloud MCP servers available and updated March 2026.
Slack, Notion, GitHub, Zapier/Make/n8n — community-built servers available. Always verify whether a server is officially supported or community-built before connecting to live business data — this matters for security review.
The MCP ecosystem now includes over 10,000 servers. The discovery directory at mcp.so lists hundreds across every category. Search there before assuming a custom build is required.
The 10-point MCP readiness checklist
Most businesses reading this will discover they are not MCP-ready — and that is not a problem. It is the point. The advantage is not in rushing to implement. It is in knowing exactly what needs fixing before you do.
How to use this checklist: Answer each question honestly. Questions 1–6 cover the foundational stages. Questions 7–10 cover MCP-specific readiness. A ‘No’ answer to any of questions 1–6 means that is where to invest first — not MCP.
1. Is your business identity consistent across all the places an AI would look for it — website, Google Business Profile, industry directories, regulatory listings?
2. Does independent, third-party evidence confirm what you claim about yourself — press coverage, directories, awards, professional body memberships, case studies?
3. Is your positioning specific enough to match a precise query — can you state exactly what you do, who you do it for, and what outcome you deliver in the language your clients use?
4. Is your key information extractable by machines — FAQ sections, schema markup implemented correctly, direct answers that machines can parse?
5. Do you know which AI systems are currently naming you — and which are not? Have you queried ChatGPT, Perplexity, and Gemini with the prompts your clients use?
6. Is your website technically sound — clean URL structure, no redirect chains, correct canonical tags, schema without errors?
7. Can you identify the specific transaction your business would enable via MCP — not “being more visible” but a concrete interaction an AI agent could complete?
8. Is your key business data structured, clean, and consistent in its source systems — not buried in PDFs or inconsistent across platforms?
9. Have you defined what the AI can and cannot do with your connected data — read, write, trigger, and off limits — documented before any code is written?
10. Does the cost and timing of MCP make commercial sense for your category right now — is the commercial case proportionate to a £5,000–£25,000+ custom build in 2026, or is waiting for platform tooling in 2027–28 the right call?
Questions 1–6 answered Yes: foundations in place — proceed to MCP assessment. Any No in 1–6: fix that first. Questions 7–10 answered Yes: you have a credible MCP business case. Any No in 7–10: resolve the gap or delay implementation.
Three entry points — mapped to business type
Entry Point A: Configure what already exists — available now
Who this suits: Shopify merchants. Businesses using HubSpot, Google Workspace, or other platforms with existing MCP server support.
What it involves: Enabling or connecting an existing MCP server to Claude Desktop. No custom development required. For Shopify specifically: your MCP endpoint is already enabled by default. Install Claude Desktop, connect it to your store, and ask buyer-style questions about your products. This is your first real-world test of how AI agents interact with your data, at zero cost.
Realistic expectation: This does not immediately change how ChatGPT.com recommends your business. It means your data is queryable by AI agents using MCP-compatible tools — primarily Claude Desktop and developer tools in 2026. The value compounds as consumer-facing MCP hosts mature.
Entry Point B: Build foundations and wait for platform tooling — 2027/28
Who this suits: Most SMBs without existing MCP-compatible platforms. Businesses in Categories Two and Three. Businesses whose checklist reveals foundational gaps.
What it involves: Fixing the foundations now so that when platform MCP tooling arrives (Shopify, HubSpot, and booking platforms are most likely in the 2027–28 window), implementation is a configuration task rather than a catch-up exercise. Your trigger is a platform announcement — not a date.
Entry Point C: Custom build — for businesses with specific operational justification
Who this suits: Category One businesses with technical teams and a specific, identified transaction to enable. The checklist must be fully green before proceeding.
Honest cost guidance: £5,000–£15,000 for a focused, well-scoped implementation. £15,000–£30,000+ for complex integrations with multiple data sources and write capabilities. The wrong reason to choose Entry Point C: “everyone is talking about MCP and we should have one.” That is not a business case.
The governance document — what to define before implementation
Every MCP implementation needs a governance document before it goes live. This is not bureaucracy — it is the difference between a controlled data layer and a security exposure. In April 2025, security researchers identified active MCP vulnerabilities including prompt injection, data exfiltration through permission combinations, and lookalike tool substitution.
1. Data scope — list every data type the server will expose. Every field should have a deliberate decision: in or out.
2. Permission boundaries — Read: what can the AI retrieve? Write: what can it create or update? (Default to read-only.) Execute: what actions can it trigger? Off limits: what is absolutely excluded — customer personal data beyond what the use case requires, internal pricing structures, regulated information.
3. Authentication and access control — OAuth 2.1 is the current standard. Which AI systems are authorised? What happens if an unknown client attempts a connection?
4. Incident response — who is notified if the server behaves unexpectedly? What is the kill switch? How quickly can access be revoked?
5. Review cadence — quarterly minimum. Review covers: is the data scope still accurate, are permission boundaries still appropriate, have new security advisories been published?
6. Regulatory and compliance considerations — does any exposed data fall under GDPR, sector-specific regulation, or contractual confidentiality? For professional services firms, a legal review should precede implementation.
Where to start this week
If you do nothing else after reading this guide, do this section.
For every business: Query ChatGPT, Perplexity, and Gemini with the prompts your clients use. Note which AI systems name you and which do not. Run the AI Recommendation Readiness Diagnostic. This tells you where you actually are before any MCP conversation is worth having.
For Shopify merchants: Install Claude Desktop (free). Connect it to your store via the native MCP endpoint. Ask buyer-style questions about your products. This costs nothing and takes under an hour.
For WooCommerce / WordPress merchants: Find and review the MCP for WooCommerce plugin. If read-only product data exposure makes sense for your business, install, configure, and test with Claude Desktop.
For HubSpot, Salesforce, or Google Workspace users: Review the official MCP documentation for your platform. Assess whether the current server covers the data you would want to expose. If you lack internal technical resource, this is a conversation for your developer or agency.
For everyone else: Complete the 10-point readiness checklist. If questions 1–6 surface any gap, that is where your resource belongs. Every hour spent on entity consistency, structured content, and external corroboration improves your AI recommendation eligibility now — regardless of MCP.
MCP readiness is not a single decision. It is a sequence: foundations first, data structure second, governance third, implementation last. The businesses that get this sequence right will find MCP is a short sprint. The businesses that skip to implementation will find it is an expensive rebuild.
This guide is part of the How to Make LLMs Recommend Your Business content programme. Read the strategic companion — MCP Will Change Which Businesses AI Recommends — for the why and whether. Take the AI Recommendation Readiness Diagnostic to identify your Stage 1–4 bottlenecks before considering Stage 5.
