MCP Servers for Non-Technical Founders: Where to Actually Start

Every non-technical founder who has heard the phrase Model Context Protocol in the last six months has followed roughly the same path. They read a thread about how MCP lets Claude plug into the tools they already use. They click through to the Anthropic docs, see a page opening with a diagram about transports and a paragraph about stdio streams, and close the tab. A week later they try again with a community list on GitHub, find three hundred entries sorted alphabetically, pick one at random, and give up when the README asks them to run a Python module. The problem is not that MCP is hard. The problem is that nothing on the first page of search results picks for them. Every resource is a catalogue or a technical spec, and a founder with sixty free minutes on a Sunday does not need either.

This is the opposite of a catalogue. There is one right answer to where to start, and it is the Google Workspace MCP server. Install it, complete a single OAuth step, and on Monday morning Claude can read your inbox, look at your calendar, and search your Drive. Three workflows land on day one and none require writing code or editing anything more technical than a small block of JSON this article hands you.

Why Google Workspace Is The First MCP Server To Install

A founder's operating system is Gmail, Google Calendar, and Google Drive, even the ones who swear by Superhuman and Notion. The inbox is where the day starts, the calendar is where attention goes, and Drive is where every deck, contract, and spreadsheet the company has ever produced eventually lives. Point Claude at those three surfaces and you cover the vast majority of questions a founder asks an assistant in a normal week. What did the investor say about the follow-on. When did I last meet the head of partnerships. Where is the pricing spreadsheet from August. A Slack MCP matters if you run a twenty-person team and a HubSpot MCP matters if you live in CRM, but every founder without exception lives in Workspace. Starting anywhere else solves a narrower problem with more setup.

The Google Workspace MCP is also one of the most actively maintained servers in the ecosystem. The OAuth flow is the one Google has polished for fifteen years, the scopes are granular, and the server exposes Gmail, Calendar, Drive, and Docs through a single connection, which means one consent screen instead of four. You can graduate later to separate servers per product if you want surgical permissions, but on day one the consolidated server is the right tradeoff between setup time and coverage.

Installing The Google Workspace MCP In Claude Desktop

The install lives entirely inside Claude Desktop, which is the macOS and Windows app that hosts MCP servers for you. Download Claude Desktop from the official site if you do not have it, log in, and confirm that the app opens normally and you can chat. That is the only prerequisite. There is no separate runtime to install, no terminal to open, and no package manager to learn. The Claude Desktop app ships with Node baked in and runs the MCP servers for you as child processes.

The configuration file is a small JSON document that tells Claude Desktop which servers to launch at startup. On macOS it lives at ~/Library/Application Support/Claude/claude_desktop_config.json and on Windows at %APPDATA%\Claude\claude_desktop_config.json. If you have never touched this file it may not exist yet. Open Claude Desktop, go to Settings, Developer, and click Edit Config. The app creates the file if it is missing and opens it in your default editor. The file is small. It has an mcpServers object, and each server is a named entry with a command and an args array. The Google Workspace MCP entry looks roughly like the following, with the exact package name taken from the server's current README because packages move and this article cannot promise the name a year from now. You paste the block into the file, save it, and quit and relaunch Claude Desktop. The quit is important. A reload from the menu does not re-read the config on every build of the app. Fully quit from the menu bar or task tray and reopen.

The first time Claude Desktop launches with a new server in the config, a browser window opens asking you to authorize the MCP to access your Google account. If the window does not open, the Claude Desktop developer panel has a logs view that will tell you whether the server failed to launch, usually because the JSON has a missing or trailing comma. If the server does launch but the browser never appears, a popup blocker is the usual culprit and copying the OAuth URL from the logs into a tab manually is the reliable fallback. Both failures are common first-install snags with thirty-second fixes.

The OAuth Step And The Read-Only Rule

The consent screen Google shows you is the single most important decision point in the whole setup, and most tutorials race through it because they are written by people comfortable with write access from day one. You should not be. When the consent screen appears it lists the scopes the server is requesting. The default Google Workspace MCP requests a broad set that includes sending email, modifying calendar events, and editing Drive files. Uncheck every scope that includes the words send, modify, write, or delete. What you want, for the first week, is read-only. Gmail readonly, calendar readonly, drive readonly, docs readonly. Those four scopes let Claude see everything and change nothing, which is exactly the posture you want while you learn what the model will and will not do with access to your inbox.

The reason for the read-only rule is not paranoia. It is that every founder who skips it has the same story. They ask Claude to clean up their inbox, the model helpfully archives a thread that looked like a newsletter but was in fact a legal notice from a vendor, and the founder spends Tuesday afternoon undoing changes. Read-only prevents the entire category. Claude can propose a reply, you copy it into Gmail yourself, and the review step happens in the interface you already know. After a week of this, when you have a sense of which requests the model handles cleanly and which it overreaches on, you revoke the OAuth token in your Google account settings, re-run the install, and consent to send and modify scopes. The second install takes two minutes. The week of read-only costs nothing and saves the first disaster.

This is the general MCP permissions rule, not specific to Workspace. Every server you add later, whether a CRM, Slack workspace, or billing platform, gets read-only access first. Write scopes come as a second install, after you have watched the model work and decided you trust what it does with the data.

Workflow One: Morning Email Triage And Summary

The first workflow that makes the install pay is the morning triage. The prompt is simple and you type it once at the start of the day. You ask Claude to read your unread emails from the last twenty-four hours, group them into three buckets, and summarize each bucket in two sentences. The buckets are things that need a reply from you today, things that are informational and require no reply, and things that are probably automated and can be ignored or archived later. Claude reads the inbox through the MCP, pulls the threads that match, and hands you back a summary that fits on one screen.

What this replaces is the twenty minutes at the start of every day where a founder scrolls the inbox, opens threads, forgets what was in them by the time they reach the bottom, and starts again. The summary collapses that into ninety seconds of reading. For the threads in the reply bucket, you ask Claude to draft a reply for each, and because you are in read-only mode the draft appears in the Claude chat rather than in Gmail. You copy the ones you like into Gmail, edit, and send. The mental model is that Claude is a chief of staff who has read everything and handed you a briefing, not an agent who is going to reply on your behalf. That distinction is what makes the first week productive rather than stressful.

The summary prompt works better if you tell Claude what counts as urgent for your role. A seed-stage founder cares about investor replies, customer escalations, and hiring pipeline. A bootstrapped founder cares about paying customers, churn signals, and partnership inbound. Spend two minutes writing a four-sentence description of what matters to you, paste it at the top of the triage prompt, and the buckets come back dramatically better.

Workflow Two: Calendar Blocking From Stated Priorities

The second workflow uses the calendar read scope. At the start of the week you tell Claude what you are trying to move forward. Three or four priorities is the right number, written as sentences not project codes. For each priority, state how many deep-work hours it would take this week to feel like it moved. Then ask Claude to look at your calendar for the week, find the gaps longer than ninety minutes that are not already blocked, and propose a set of focus blocks that cover your priorities. Claude cannot create the events because you are in read-only mode, but it can give you a list of times to block, and you click through to the calendar and create the holds yourself in under two minutes.

What this replaces is the planning habit nobody actually does. Every productivity book tells the founder to time-block the week on Sunday evening. Most founders do it for two weeks and stop because the exercise requires remembering priorities, scanning the calendar, and doing the spatial Tetris of fitting blocks into gaps at the same time. Claude handles the last two and the founder supplies the first. That is the right division of labor.

The prompt improves with a constraint list. Tell Claude you do not take focus blocks before 10am, that Fridays are for customer calls, and that a three-hour priority is better as one long block than two short ones. These constraints stack up over a month and eventually you have a small Monday prompt that reflects how you actually work.

Workflow Three: Natural-Language Drive Search

The third workflow unlocks the archive. Every founder has a Drive with four hundred files where the search box only finds documents whose exact title you can remember. Claude with Drive read access can search by meaning. Find the deck I sent the lead investor in March. Show me the last version of the pricing spreadsheet. Where is the customer contract template we used for the enterprise deal last quarter. The MCP hands Claude the ability to search files and read their contents, and the model matches on context in a way the native Drive search cannot.

Founders underestimate this workflow until they use it. The first time you find a document in six seconds that would have taken fifteen minutes of scrolling, the install has paid for itself. A useful habit is to start meeting prep by asking Claude to gather relevant documents. Before an investor update, ask for the last three board decks and the most recent financial model and you are briefed in under a minute. The value is less about any single document and more about the collapse of the cost of retrieval.

Prompting Claude Like An Executive Assistant

The three workflows above work, but they work dramatically better when the prompts are written in a specific style. The pattern is simple. Treat Claude as a new executive assistant on day one. A new assistant is smart and motivated but does not yet know your context, your preferences, or the constraints you operate under. You would not ask a new assistant to clean up your inbox without first telling them what counts as urgent, which senders always get a reply, and what tone you write in. You would not ask them to block focus time without telling them when you think best and what meetings are immovable. Every productive prompt in this system has three parts. The context you bring, the constraints that apply, and the specific request. When the prompt is missing context the output feels generic. When it is missing constraints the model over-reaches. When it is missing a specific request the answer is a list of options rather than a decision.

The second rule, which is really a corollary of the read-only posture, is that every write action gets a preview step. Even after you graduate to send and modify scopes, never tell Claude to send an email. Tell it to draft an email and show you the draft, confirm, and then send. The preview costs two seconds and prevents every class of error that makes founders abandon AI assistants after a bad week. The assistants who last are the ones who ask before acting.

Where Founders Get Stuck And What To Do Next

The sequence above is twenty minutes for someone who has installed an MCP server before and a lost afternoon for someone who has not. Every step is documented, but the documentation is scattered across three sites and two communities, and the failure modes are the kind where knowing the fix is ninety seconds and finding it is an hour. If the OAuth window does not appear, if the Workspace admin in your company has blocked third-party app consent, if the config JSON parses but Claude still does not see the server, you will spend more time Googling than building. That is the real cost of getting started with MCP, and it is why the founders who land it do so either on a weekend with time to debug or with someone who has done the install enough times to know the sharp edges. The WitsCode team runs guided MCP onboarding for founders who would rather spend Sunday on their roadmap than on Claude Desktop logs, and the install arrives complete with the three prompts above tuned to your operating style. The server works for everyone. The prompts are what make it worth keeping.