MCP guides

MCP guides · Founders

MCP is how Claude stops being a chat window and starts being an agent team.

These are the MCP servers a founder hooks up first: the ones that turn Claude from a smart text generator into a thing that actually moves your week forward.

Why this matters

Model Context Protocol is the most consequential thing Anthropic has shipped since tool use. It lets a Claude conversation reach into your CRM, calendar, file store, and codebase as a first-class participant instead of a copy-paste middleman.

The catch: most MCP content right now is either Anthropic's reference docs or someone's hello-world server. There is almost no honest writing about which MCP servers actually pay off for a founder running the company in 30 tabs and a Slack drafts folder.

This guide is the version a working founder needed six months ago: the MCP servers that earn their keep, the ones that look impressive but do not, and how Nerve composes the worthwhile ones into an agent team that closes loops on its own.

What MCP unlocks here

Gmail MCP: the inbox is where promises die

An MCP server with read + search + draft over Gmail turns Claude into a thing that can triage your inbox, surface promises you made to customers and investors, and draft the replies that have been sitting in your head for nine days. Most founders should wire this up first.

Google Calendar MCP: actual time, not vibes

Calendar MCP lets Claude propose meeting times that are real, book holds, and reason about your week. Without it, every 'find me 30 minutes' devolves into manual screenshots. Hugely underrated as a founder primitive.

Salesforce / HubSpot MCP: kill the swivel-chair

CRM MCP servers let an agent push pipeline updates, log activities, and surface deals going stale, without you switching tabs. Salesforce's official MCP work and HubSpot's API surface both have community wrappers worth trying. Nerve targets the writes that matter (next-step, stage, owner) over the breadth.

Slack MCP: the surface where work actually happens

Slack MCP is a quiet workhorse. Read channels and threads, search, send DMs as you. Combined with a CRM connector, an agent can ask 'did we follow up with X' and answer it from the real evidence, not from memory.

Filesystem + Notion MCP: where the docs live

Most founder context is in unstructured docs: pitch deck v17, meeting notes, an old proposal. MCP servers over your file store let an agent retrieve and cite the right paragraph. Nerve's gdrive MCP tool was built for exactly this.

GitHub MCP: only if you ship code

If you are a technical founder still in the codebase, GitHub MCP is a fast win: PRs, issues, code search, commit drafting. If you've moved to managing builders, it's lower priority than CRM and inbox.

Three steps to a working agent

1

Pick a host that speaks MCP

Claude Desktop and Claude Code both ship MCP. Cowork agents and OpenAI's MCP support are catching up. The Nerve MCP endpoint at app/mcp/route.js is what we use for our own agents; sign up and you get a key that any MCP host can hit.

2

Wire 2 to 3 servers, not 12

Start with Gmail, Calendar, and your CRM. Resist the urge to plug in everything on day one. Each new server adds tool noise. The 80% of value lives in the 20% of servers that touch where your actual day happens.

3

Give the agent a job, not a chat

MCP unlocks its real value when you stop chatting and start asking the agent to own a recurring outcome: 'tell me every morning what changed in my pipeline,' 'draft replies for every overdue customer thread,' 'book the three meetings I owe.' That's where Nerve's agent team picks up.

What people doing this say

The first time the agent actually moved a deal forward without me copy-pasting anything, I stopped treating Claude as a chatbot. MCP is the wire.

Solo founder, B2B SaaS (paraphrased)

We tried four CRM AI tools and one MCP server. The MCP server beat all four because it wrote to the right field instead of guessing.

Operator, 12-person seed-stage startup

MCP is the difference between 'AI that helps me think' and 'AI that helps me ship.' Founders need the second one.

Patrick Hillstrom, Nerve

Common questions

What is MCP, briefly?

Model Context Protocol is Anthropic's open spec for letting LLMs call external tools through a standard JSON-RPC interface. A server exposes tools (read_emails, create_event), a host (Claude, Cowork, Nerve) calls them. It is the cleanest way to give an LLM real-world reach without bespoke integrations per tool.

Are MCP servers safe to install?

Treat MCP servers like browser extensions. Read the source if it is community-maintained. Prefer servers from the company that owns the underlying data (Salesforce's official MCP work, for example) when you can. Scope auth tokens to the minimum permissions the server actually needs.

Do I need to build my own MCP server?

Usually no. The community-maintained servers cover Gmail, Calendar, Slack, GitHub, Linear, Notion, filesystem, and most major CRMs. Build your own when the data you care about is proprietary (your own product database, an internal tool, a stack the community has not gotten to yet).

How does Nerve relate to MCP?

Nerve is MCP-native on both ends. The Nerve agents (briefing, pipeline, fundraising, customer success) talk to your stack over MCP. The Nerve MCP server lets external hosts (Claude Desktop, Claude Code, Cowork) query and act on your Nerve state. Patrick built it this way because composing existing servers beats reinventing them.

Will MCP get standardized further?

OpenAI and other LLM vendors are converging on MCP. The 2025 to 2026 trajectory is MCP becoming the default for tool calling across model providers. Investing in MCP-shaped tooling now is a safe bet.

Roles where this matters most

Stop chatting with Claude. Start running an agent team.

Start free, no credit card

Other MCP guides