MCP guides · Setup

MCP is simpler than the docs make it look. Here's the 30-minute version.

If you have an MCP host (Claude Desktop, Claude Code, or Cowork) and one MCP server, you have a working agent. The hard part is choosing which server to install first.

Try Nerve free Read Nerve's developer docs

The problem

Why this matters

Most MCP intros open with the protocol spec, JSON-RPC, capability negotiation, and 14 lines of YAML configuration. By the time you finish reading, you have not actually installed anything and you have lost an hour.

The reality is much simpler. MCP is a way to give an LLM tools through a standard interface. There is a host (where the LLM runs), there are servers (where the tools live), and there is a config file that tells the host which servers to talk to. That's it.

This guide is the version that gets you from zero to a working agent before lunch: which host to use, which server to install first, how to handle auth without losing a day, and what to do once your hello-world works.

What you get

What MCP unlocks here

Pick a host that does most of the work

Claude Desktop is the lowest-friction starting point: a config file, a server name, a path. Claude Code is identical for developers. Cowork runs as a hosted agent. Skip OpenAI for now if you are MCP-first.

Start with one server, not seven

Gmail OR Calendar OR Filesystem. Pick the one whose data is the bottleneck in your day. Get it working end-to-end before adding more. Tool noise is the silent killer of MCP setups.

Auth is the hard part, accept it

OAuth is where you'll spend most of the time. Read the server's README before installing. Strip scopes down to least privilege. Workspace admin policies will sometimes block scopes; ask your admin or use a personal account for the prototype.

Make the agent do a recurring outcome on day one

Once the server is wired, do not just chat. Ask the agent for an actual deliverable: 'tell me what's overdue,' 'find me 30 minutes with X,' 'summarize what changed in folder Y since Tuesday.' Recurring outcomes are the test of whether the install was worth it.

Setup

Three steps to a working agent

Choose your host

Claude Desktop for solo prototyping. Claude Code if you live in the terminal. Cowork if you want a hosted agent that runs without your laptop being open. All three speak the same MCP. Hosts are interchangeable; servers are reusable.

Install one server, end-to-end

Clone the server repo. Read the README for environment variables. Get OAuth working (this is the long part). Add the server to your host's config. Restart the host. Verify the tools show up. Verify a real call works (read one email, list five calendar events, list directory contents).

Give the agent a job

Stop chatting. Start asking for outcomes. 'Every morning, list the threads I have not replied to.' 'Find time with three people this week.' 'Summarize the doc folder for last month.' When the recurring outcome works, you are done. Add servers from there.

FAQ

Common questions

Do I need to be a developer to set this up?

For Claude Desktop with a single server, no, you can copy a JSON config and follow a README. For more complex setups (multiple servers, custom auth, self-hosted servers), some familiarity with environment variables and OAuth helps. Cowork hides most of the plumbing behind a web UI.

What's the difference between an MCP host and an MCP server?

Host = the thing the LLM runs in (Claude Desktop, Claude Code, Cowork). Server = the thing that exposes tools (Gmail server, Calendar server, your custom server). Hosts call servers over a standard protocol. One host can talk to many servers.

Are MCP servers safe to install?

Treat them like browser extensions. Prefer official servers from the data provider when they exist. Read community server source code. Scope auth tokens to minimum permissions. Audit logs if the server provides them.

Can I build my own MCP server?

Yes. The SDK is straightforward in TypeScript and Python. Most founders only need to build their own when the data is proprietary (an internal database, a custom tool). For Gmail, Calendar, GitHub, Linear, Notion, and major CRMs, the community already shipped one worth using.

How does Nerve fit in?

Nerve is an MCP server, not a host. It exposes tools like nerve_briefing and nerve_find_time that Claude Desktop, Claude.ai, or Claude Code can call: read-only on the Pro plan, read and write on Operator. Connect it and your LLM can work with your Nerve briefing, follow-ups, and calendar.

Stop reading MCP docs. Install one server and feel the unlock.

Start free, no credit card

More MCP guides