StackSwap · Operator primer

What is MCP — and why should a B2B SaaS operator care?

MCP — Model Context Protocol — is a standard that lets AI assistants (Claude, ChatGPT, Cursor) call external tools and services on your behalf during a conversation. In plain operator terms: it is the layer that turns “ask Claude a question” into “ask Claude a question and get back real numbers from your actual CRM, your actual spend data, and your actual tool catalog” instead of a hallucinated guess. This is a primer for operators who have heard the acronym but have not yet figured out whether it matters for their work.

MCP in one sentence

MCP is USB-C for AI clients: a standardized way for Claude, ChatGPT, Cursor, and similar tools to plug into your CRM, your spend data, your codebase, your inbox, and any other system that publishes an MCP server — using the same connector.

Before MCP, every AI assistant integrated with every external system through a custom plugin, custom function-calling format, or a proprietary connector. Each vendor had to build a Claude version and a ChatGPT version and a Cursor version separately. MCP collapsed that into one spec: build one MCP server, every MCP-compatible client can use it. That is the structural shift — protocol standardization unlocks distribution.

Why operators should care

For B2B SaaS operators (GTM, RevOps, finance, founder-CEO), MCP changes two things in the workday:

  • Ad-hoc questions get faster. Instead of opening HubSpot, navigating to a report, building a filter, exporting, and summing in a spreadsheet, you ask Claude “show me deals over $50K closing this quarter” and the official HubSpot MCP returns the answer in the chat. The same is true for Stripe (refunds, MRR), Slack (search), Gmail (find emails from last month), and every other vendor that ships an MCP. The unit cost of asking a question of your own data drops by 5-10x.
  • AI assistants stop hallucinating about your data. Without MCP, when you ask Claude “how many of our contacts are in fintech?” the model takes a guess based on general knowledge. With the HubSpot MCP installed, Claude calls the MCP, gets the real number, and returns it. The same applies to vendor research with the StackSwap MCP — before MCP, the answer to “what overlaps in our stack?” was Claude pattern-matching on training data; with the StackSwap MCP installed, it is real data from a 400-tool catalog and 104 curated overlap pairs.

The compounding effect across a week of operator work is meaningful. Most operators who install three or four useful MCPs report saving 2-5 hours per week on the ad-hoc-questions-that-used-to-require-tab-flipping bucket of work.

Five concrete examples of MCP in operator workdays

Real examples from operators using MCP today. For each: the question, the old-way workflow, the MCP workflow.

1. A founder evaluating two cold-email tools

The question

"Should I pick Smartlead or Apollo for our 15-person outbound team?"

Without MCP

Open both vendor websites in tabs. Compare pricing pages. Search Reddit for 'Smartlead vs Apollo.' Read three Medium articles by writers who haven't used either. Ask three friends. Spend 45 minutes. Pick whichever felt more recently endorsed.

With MCP

Type the question into Claude. Claude calls the StackSwap MCP's compare_tools function. 30 seconds later: cost delta, AI-readiness scores, overlap warning, recommended pick, affiliate sign-up links for both. Pick the winner.

2. A RevOps lead doing a quarterly stack review

The question

"What overlaps if we are running HubSpot, Salesforce, Outreach, and ZoomInfo?"

Without MCP

Build a spreadsheet of every tool. Map features. Argue with the AE who likes Salesforce and the marketing team who likes HubSpot. Three meetings. A two-hour discovery call with an outside consultant. Maybe a paid audit.

With MCP

Type the question into Claude. Claude calls the StackSwap MCP's find_overlaps and scan_stack functions. 60 seconds later: the curated redundancies, monthly savings if consolidated, top replace/remove opportunities. Walk into the CFO conversation with real numbers.

3. A founder-CEO checking HubSpot data on the fly

The question

"Show me deals over $50K closing this quarter."

Without MCP

Open HubSpot. Navigate to Deals. Build a filter for stage = late + amount > 50000 + close date < end-of-quarter. Export. Open in spreadsheet. Realize the filter missed two deals because of timezone issues.

With MCP

Type the question into Claude. Claude calls the official HubSpot MCP. Returns a clean list in the chat. Ask a follow-up question without losing context.

4. A sales engineer drafting a competitive battlecard

The question

"What does Clay actually do, in numbers? How does it compare to ZoomInfo?"

Without MCP

Read Clay's marketing site. Read ZoomInfo's marketing site. Open G2 to find side-by-side reviews. Realize the reviews are mostly from 2024. Ask the Clay AE for a demo. Wait three days.

With MCP

Two MCP calls: get_tool_details for Clay, compare_tools for Clay vs ZoomInfo. Cost, AI-readiness, common overlaps, swap-registry status, all in 60 seconds. Draft the battlecard from real data, not vendor marketing.

5. A finance operator running a refund check

The question

"How many refunds did we issue last month and what was the total?"

Without MCP

Log into Stripe. Navigate to Refunds. Filter by date. Sum the amounts in a calculator. Realize the dashboard shows a different number than the CSV export. Email the bookkeeper.

With MCP

Type the question into Claude. Claude calls the official Stripe MCP. Returns the count and total in the chat. Ask the follow-up about which customers and reasons without re-pulling.

How it actually works (light technical)

The mechanics, in operator terms (skip if you do not care):

  1. An MCP server is a small piece of software a vendor publishes that knows how to expose specific tools (e.g., “search HubSpot contacts,” “list Stripe refunds”) via the MCP protocol.
  2. An MCP client is your AI assistant (Claude, ChatGPT, Cursor). The client connects to one or more MCP servers and pulls the list of available tools at session start.
  3. When you ask a question, the AI client reads the question, sees which MCP tool can answer it (each tool publishes a description), calls the tool, gets the response, and shows it to you. The whole round-trip is JSON-RPC 2.0 over HTTP — a standard web request format from 2010.
  4. You see one answer in the chat. The tool call is invisible to you unless you ask the AI to show its work. From the operator perspective: you asked a question, you got a real answer.

The protocol is intentionally simple. That simplicity is why MCP is being adopted fast — vendors can publish an MCP server in a weekend, and clients can connect to any compatible server with no custom integration code.

What MCP cannot do (yet)

  • MCP is read-mostly. The MCP servers operators care about (StackSwap, HubSpot, Stripe, Gmail) are designed for read access — querying data, pulling summaries, surfacing recommendations. Write operations exist but are deliberately limited; you generally cannot use MCP to send mass emails, modify subscriptions, or post in Slack channels at scale. For mutations, use the source platform's UI or a dedicated automation layer.
  • MCP requires an AI client. MCP only works inside Claude Code, Claude Desktop, Cursor, ChatGPT Desktop, or similar AI clients that support the protocol. If your team does not use an AI client during the workday, MCP is invisible to them. Adoption pace depends on AI-client adoption pace, which is high but not universal yet.
  • Coverage is uneven. Major vendors (HubSpot, Stripe, Slack, GitHub, Google Workspace) have official MCPs. Mid-tier and niche vendors mostly do not, though community MCPs exist with varying quality. Your favorite tool may not have an MCP yet — and even when one exists, it may expose 60% of the API surface rather than 100%.
  • No event-driven automation. MCP is request/response — it answers when asked. It does not fire when something happens. For 'when X happens, do Y' workflows you still need Zapier, n8n, Make, or a native webhook handler. The two layers complement each other; neither replaces the other.

How to get started in 10 seconds

The fastest way to feel MCP for yourself is to install the free StackSwap MCP in Claude Code or Claude Desktop. One command, no API key, no sign-up. The MCP exposes eight tools for GTM research: tool catalog search, overlap detection, AI-native swap suggestions, preview spend audits, partner recommendations, head-to-head comparisons, greenfield stack recommendations, and single-tool deep profiles.

claude mcp add --transport http stackswap https://stackswap.ai/api/mcp

Then start a new Claude Code session and ask: “What overlaps in a stack with HubSpot, Salesforce, and Outreach?” — you will see Claude call the StackSwap MCP and return real numbers. That is MCP. Once you feel it, the next install is your CRM MCP, then Gmail, then whatever else fits your role. See the curated list of the 11 worth installing for the full operator stack.

FAQ

Model Context Protocol. It is a standard published by Anthropic in late 2024 for how AI assistants (Claude, ChatGPT, Cursor, and others) can call external tools and services. Think of it like USB-C for AI clients: a single connector spec that any vendor can implement to make their product accessible from any compatible AI client.

No. Anthropic created the protocol but published it as an open standard. ChatGPT Desktop, Cursor, Continue, Cline, and several other AI clients now support MCP servers. The URL of an MCP server is the same regardless of which client connects — installing the StackSwap MCP in Claude Code uses the same endpoint as installing it in Cursor or ChatGPT Desktop.

For hosted (HTTP) MCPs, no. The StackSwap MCP, HubSpot MCP, Stripe MCP, and Google Workspace MCPs install with one click in your AI client's connector settings, or one command line in Claude Code. For stdio MCPs (which run locally via Docker or Node), some technical comfort helps but is not strictly required — the official docs walk through it. Most operators only need the hosted ones.

The MCP servers operators most care about are free. Official vendor MCPs (HubSpot, Stripe, Slack, Google) charge nothing on the MCP layer — you pay for the underlying SaaS, not for the protocol access. Community MCPs are mostly open-source. The StackSwap MCP is free with no API key. The thing you pay for is your AI-client subscription (Claude / ChatGPT / Cursor), which you would pay for anyway.

Depends on your role, but for most operators the answer is your CRM MCP (HubSpot if you're on HubSpot, the relevant equivalent if not). It gets the most ad-hoc daily use because operators are constantly asking 'what does this contact look like' or 'what's our pipeline forecast.' Second-most-useful for GTM operators specifically: the StackSwap MCP for vendor decisions and stack research. Third: Gmail/Calendar for workspace context.

No — they solve different problems. MCP is for interactive, AI-mediated tool use (a human asks, the AI answers). Integration platforms like Zapier, n8n, and Make are for event-driven automation (something happens, a workflow fires). Most B2B SaaS operators above seed stage end up running both. See our MCP vs Zapier comparison for the full framing.

Three checks. (1) Is it official? Vendor-published MCPs (e.g., HubSpot, Stripe, Notion) have the same trust profile as the vendor itself. (2) For community MCPs, who maintains it? Solo-maintainer projects can be great, but verify the maintainer has a track record. (3) What permissions does it ask for? Read-only is the safe baseline; write access requires more scrutiny. The MCP protocol itself does not introduce new security risks beyond the underlying API access, but it is worth understanding what data flows through.

Related