An MCP server is the piece that lets an AI agent use a tool on its own. ZeroCrew runs one at mcp.zerocrew.ai, so this guide explains MCP servers with a real one you can connect today: what the server offers, what a request looks like, and what comes back.
Founders ask what an MCP server is because the agents they already pay for, like Claude, ChatGPT, and Cursor, can do far more once a server is connected. ZeroCrew's MCP servers list shows 104 business tools with an MCP server across 11 roles, as of November 2026, so the answer matters for almost every job in a one-person company.
What is an MCP server?
An MCP server is a program that gives an AI app context and actions from one tool, through the Model Context Protocol. The official MCP architecture docs define it as "a program that provides context to MCP clients". ZeroCrew's own MCP server is a small, live case: it lets your agent search the ZeroCrew directory of AI agents and tools.
The MCP docs name three parts in every connection, and ZeroCrew's server shows where each part sits:
- The host is the AI app you work in, like Claude, ChatGPT, or Cursor.
- The client is the connector the host creates for each server. One host keeps one client per server.
- The server is the tool's side. ZeroCrew's server answers for the ZeroCrew directory; Stripe's server answers for your payments.
The Model Context Protocol itself is an open standard, and ZeroCrew's guide What is MCP? covers the protocol in plain words. Anthropic released MCP in November 2024, and on December 9, 2025 the Linux Foundation announced the Agentic AI Foundation as MCP's new, neutral home, with Anthropic, OpenAI, Google, Microsoft, and others as members.
What does an MCP server offer an AI agent?
An MCP server offers three kinds of building blocks: tools the model calls, resources the app reads, and prompts the user picks. The MCP server concepts docs describe all three, and most servers on ZeroCrew's MCP servers list lead with tools, because tools are what let an agent do the work.
| Building block | What it is | Who controls it | Example |
|---|---|---|---|
| Tools | Functions the model calls | The model | Search flights, send a message |
| Resources | Read-only data for context | The app | A file, a database schema |
| Prompts | Ready-made task templates | The user | Summarize my meetings |
Each MCP tool comes with a name, a plain-language description, and a schema for its inputs, so the agent reads the menu and decides when to call it. ZeroCrew's server offers four tools and no resources or prompts: a few clear actions built around one job, searching the directory.
How does an MCP server work, step by step?
An MCP server works in five steps: the app connects, lists the tools, calls one, gets results, and goes deeper. The MCP docs name the two key requests tools/list and tools/call, and ZeroCrew's server answers both. Here is the exchange when an agent connects to ZeroCrew:
- The app connects. ZeroCrew's server introduces itself as zerocrew-directory and says what it's for: search the directory of AI agents, apps, and models that founders use to run a company.
- The app lists the tools. The server answers with four: search_tools, get_tool, list_roles, and list_job, each with a description and its inputs.
- The model picks a tool. You ask "which AI tools can answer phone calls for my business?", and the model calls search_tools with that sentence as the query.
- The server answers. The reply holds up to 20 tools, best match first. On November 3, 2026, the first results were Bland AI, Retell AI, Rosie, Dialzara, Goodcall, and Synthflow, each with its job and its ZeroCrew page.
- The agent goes deeper. The model calls get_tool on one result for what it does, its pricing, its MCP server address, and its alternatives, then writes you an answer in plain words.
Nobody wrote a custom integration for an agent to search ZeroCrew's directory, and that is the point of an MCP server. ZeroCrew described its directory once, in four tools, and every MCP app can now use it the same way.
What is the difference between a local and a remote MCP server?
A local MCP server runs on your own computer and talks over stdio, while a remote MCP server runs on the internet and talks over Streamable HTTP. The MCP architecture docs describe both transports, and ZeroCrew's server is a remote one: you add one web address and nothing gets installed.
| Local MCP server | Remote MCP server | |
|---|---|---|
| Where it runs | On your machine | On the tool's servers |
| Transport | stdio | Streamable HTTP |
| Serves | Usually one client | Many clients at once |
| Good for | Files, local apps, scripts | Business tools like CRMs, payments, and directories |
According to the MCP docs, a local server that uses stdio typically serves a single client, while a remote server typically serves many. For a founder, most of the business tools on ZeroCrew's MCP servers list run remote servers, because your agent can reach them from any device or from a cloud agent.
How is an MCP server different from an API?
An API is built for programmers who read the docs and write code, while an MCP server is built for AI agents that read a tool's menu and use it on their own. Many MCP servers run on top of a tool's existing API. ZeroCrew's guide What is MCP? compares the two in more detail for founders who don't code.
The MCP specification says it takes inspiration from the Language Server Protocol, which lets every code editor support a programming language through one shared protocol. ZeroCrew sees the same effect in its directory: once a tool like HubSpot or Attio ships an MCP server, Claude, ChatGPT, Cursor, and any other MCP app can use it without a custom integration for each one.
Which AI apps can connect to an MCP server?
Claude, ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code all connect to MCP servers, according to the Linux Foundation's December 2025 announcement, which also counted more than 10,000 published MCP servers. ZeroCrew lists the main AI apps in its directory, so you can compare them before picking where your agents run.
ZeroCrew sorts the ways to connect an MCP server into three: a chat app, a coding agent, or an automation tool, and the choice depends on the work:
- A chat app for everyday business work. Pick Claude or ChatGPT when you want to ask questions and approve actions in a conversation.
- A coding agent for building. Pick Claude Code or Cursor when the agent also writes and runs code. ZeroCrew's coding agents page ranks the options three ways.
- An automation tool for jobs on a schedule. Pick Zapier or n8n when the work runs without you. Both have MCP servers, and ZeroCrew's automation page compares them with Make.
Is an MCP server safe to connect?
An MCP server is as safe as the access you give it and the company that runs it. The MCP specification asks AI apps to get your consent before any tool runs, and treats a tool's description of itself as untrusted unless the server is trusted. ZeroCrew marks all four of its tools read-only.
Researchers take MCP server security seriously. A study of MCP by Hou and colleagues (arXiv, revised October 2025) maps 16 threat scenarios across four kinds of attacker: malicious developers, external attackers, malicious users, and security flaws, and proposes safeguards for each stage of a server's life. ZeroCrew's advice follows from that research: start with read-only access, approve actions before they run, and connect servers only from companies you trust.
The safest order to give an agent access through MCP, the order ZeroCrew recommends, is step by step:
- Read. Connect a server that only looks things up, like ZeroCrew's directory server, or give a tool like Stripe or HubSpot read-only access first.
- Draft. Let the agent prepare emails or invoices that you approve before anything goes out.
- Act. Once you've watched the agent work, allow small actions that are easy to undo.
How do you connect ZeroCrew's MCP server to your agent?
You connect ZeroCrew's MCP server by adding mcp.zerocrew.ai as a custom connector in Claude, Cursor, or ChatGPT, with no sign-in and no API key. Your agent can then search ZeroCrew's directory for the right tool for any job, read a tool's page, and list the tools in each role, before you spend a day comparing.
- Copy the address. ZeroCrew's MCP servers list shows the full address at the top with a Copy button.
- Add it to your AI app. Open the app's connector settings and add a custom connector with that address.
- Ask a real question. Try "which bookkeeping tools have an MCP server?" or "what should I use for cold email?".
- Follow the links. Every answer from ZeroCrew's server links to the tool's page on zerocrew.ai, where the job page shows the alternatives side by side.
ZeroCrew also helps with the next step, once your agent knows which tools to use. The MCP servers list groups every tool with an MCP server by role, so you see which of your tools your agents can reach today, from Supabase and Vercel in engineering to Notion in operations. ZeroCrew's engineering page sorts 27 coding agents and tools into four jobs.
If you'd rather build an MCP server for your own product, ZeroCrew's guide How to build an MCP server walks through it step by step. If you'd rather have the agents set up for you, ZeroCrew's AI agent setup picks the agents for each job and connects them to the tools your business already uses.
What can one founder do with MCP servers?
One founder with a few MCP servers connected can hand whole jobs to agents: payments, support, outreach, and research. ZeroCrew exists for that founder, building big with zero employees, and the ZeroCrew directory is where you find the next tool for your agents to use. Connect one server today, ask it one real question, and add the next.