MCP explained for growth teams (no PhD required)

Every growth team you know is already using Claude or ChatGPT for half their daily work. Writing emails, drafting landing pages, cleaning lists, pulling insights out of reports. The part that still hurts is the data step. You paste a CSV into the chat, you ask Claude to enrich it, Claude hallucinates half the fields because it cannot reach the live web, and you paste the result back into a spreadsheet.

MCP — the Model Context Protocol — is what fixes that step. It is a small open standard (shipped by Anthropic in late 2024, now supported by Claude Desktop, Cursor, Claude Code, Zed, and a dozen others) that lets a chat client call external tools during a conversation. When you install the Stekpad MCP server, Claude gains five new verbs: scrape, crawl, map, extract, search. It can now actually fetch pages, enrich rows, and return real data — not a guess at what the data probably looks like.

This post explains MCP in plain language for a growth team. No protocol diagrams. No "tool-calling" jargon. Just what it is, what changes in your daily workflow, and how to install it in under a minute.

What MCP actually is

Think of MCP the way you think of a USB port. Before USB, every peripheral had its own cable and its own driver. After USB, you plug in any device and the computer knows how to talk to it. MCP does the same thing for AI chat clients. Before MCP, every tool integration had to be hard-coded into Claude. After MCP, any server that speaks the protocol can be installed into Claude and the chat knows how to call it.

The protocol defines three things:

1. A list of tools the server exposes, each with a name, a description, and a JSON schema for its inputs.
2. A way for the client (Claude) to call a tool with arguments and receive a response.
3. A way for the server to return resources (files, datasets, live data) that Claude can read.

That is basically it. MCP is not a new AI model, a new agent framework, or a new standard for prompting. It is a plug for tools, agreed on by the big chat clients, so you can install scraping, browsing, GitHub, Notion, Postgres, Gmail, Linear and Stekpad with one config file instead of five bespoke integrations.

Why growth teams should care

You already know the symptom. You open Claude, you ask "give me the HQ location and email pattern of these 20 companies", and Claude gives you a plausible-sounding answer for each. Some of the answers are right. Some of them are wrong. You have no way to tell which, because Claude does not have live web access unless you bolt it on.

MCP is how you bolt it on. Once you install the Stekpad MCP server, the same prompt produces a different sequence:

• Claude reads "enrich these 20 companies"
• Claude calls the scrape tool on each company URL
• Claude calls the extract tool with a schema for HQ location and email pattern
• Claude waits for the run to finish
• Claude formats the results back into your chat as a clean table

You see the calls as they happen in the chat — each tool invocation shows up as a collapsed block you can expand. You see the run_id. You see the credit cost. You can click through to the dataset in the Stekpad web app and confirm every row.

The difference is not cosmetic. It is the difference between an LLM that guesses and an LLM that looks things up.

An MCP-enabled chat is the smallest possible agent. You describe the goal in English, the chat picks tools, calls them in order, and returns real data. No workflow builder, no canvas, no glue code.

A real example: enriching 20 companies

Say a sales person has 20 company names and wants HQ location, employee count, and the email pattern. The old workflow is an afternoon: open each site, scrape the about page by hand, search LinkedIn, guess the email format, paste into a sheet. The MCP workflow is one message:

Claude plans the work, calls search to find each website, calls extract with a schema on the top result, and comes back with a table. It takes about 90 seconds and costs roughly 200 credits on the Builder pack — about 60 cents.

The dataset is now in your Stekpad workspace. You can open the web app, sort by employee count, filter by country, export to CSV. You can run the same prompt next quarter and the dataset updates in place. You can share the dataset with a teammate on Cloud Growth. The chat was the entry point, but the data lives in a dataset, not in the transcript.

Install the Stekpad MCP server in Claude Desktop

Three steps, under a minute.

Step 1: get your API key. Sign in to Stekpad, go to Settings → API Keys, click Generate. Copy the key (starts with stkpd_live_).

Step 2: open Claude Desktop's MCP config. On macOS, the config lives at ~/Library/Application Support/Claude/claude_desktop_config.json. On Windows, at %APPDATA%/Claude/claude_desktop_config.json. Create the file if it does not exist.

Step 3: paste this block. Replace stkpd_live_... with your real key.

Quit and reopen Claude Desktop. In a new chat, look at the tool icon at the bottom of the input — you should see stekpad listed with five tools: scrape, crawl, map, extract, search. Type a prompt that needs live data and watch Claude pick the right tool. If you get stuck, the full MCP install guide has screenshots and troubleshooting for every platform.

Cursor and Claude Code

Same JSON block, different file path. Cursor reads from ~/.cursor/mcp.json. Claude Code reads from ~/.config/claude-code/mcp.json. Both support the same Stekpad server. See the Cursor integration and the Claude Code integration for the specifics.

MCP vs Zapier vs n8n: the right mental model

Growth teams often ask "is MCP just Zapier for Claude?" It is not, and the difference matters.

Zapier and n8n are pre-wired flows. You build a graph: trigger, condition, action, action, action. It runs the same way every time. You can tweak the parameters, but the shape is fixed. This is great when you know the shape of the problem in advance — "new Typeform submission → enrich → add to HubSpot". It is bad when the shape changes every time you run it.

MCP is agent-driven dynamic calls. There is no graph. Claude decides which tools to call, in what order, based on your prompt and the intermediate results. If the first search returns a weird domain, Claude will go find another one. If the extract fails on one URL out of 20, Claude will tell you and move on. You describe the goal once, the agent figures out the steps.

Neither is better in the abstract. If you run the same 50,000 enrichments a month on the same schema, Zapier or a cron-triggered script is probably the right tool (and you can drive Stekpad from Zapier too). If your growth team runs a different enrichment every week — new company list, new fields, new criteria — MCP is the shape that fits. See the Zapier replacement manifesto for the deeper argument.

Pre-wired flows for repeatable work. Agent-driven MCP calls for exploratory work. Both live in your stack. They are not rivals.

What you can do with the Stekpad MCP server today

Five tools, a handful of shapes. Here are the prompts we actually run most often on our own team:

• "Scrape https://example.com and give me the contact email, the company description, and the tech stack."
• "Crawl the /blog section of https://competitor.com, store it as a markdown bundle, and give me a five-bullet summary of the last ten posts."
• "Search Google for 'best tools for invoicing EU freelancers', scrape the top ten results, and extract product name, pricing, and main feature from each landing page."
• "Map all URLs on https://docs.stripe.com, filter to /api/\\, and give me the count per top-level section."
• "Extract founder name and founding year from these 50 Crunchbase URLs and append to the 'Founders research' dataset."

Each of those runs in under two minutes in a single Claude Desktop message. Each uses Stekpad under the hood. None of them required a new script, a new Zap, or a new hire.

Safety and auditability

One honest note. When you install any MCP server, you are giving the chat client permission to call tools on your behalf. You want to know what it called and why. Two things help.

The run log. Every tool call Stekpad makes on your behalf is logged in your workspace as a Run. You can open the web app and see the full list: timestamp, verb, parameters, credits charged, status. Nothing is hidden behind the chat transcript.

The credit cap. On the free plan you get 300 credits a month, which is enough to stop an agent loop from running away. On paid plans you set a monthly hard cap in Settings → Budget. If an agent decides to crawl a 50,000-page site, you want the cap to catch it before it catches you.

Next steps

• Read the full MCP server reference for every tool description, every input schema, and every error code.
• Follow the Claude Desktop integration guide for screenshots and platform-specific quirks.
• Read the enrichment pipeline with Claude guide for a longer worked example that chains extract and 4 enrichers into a reusable pipeline.
• If you are on Cursor or Claude Code, the Cursor integration covers the same setup in the editor.