What are you actually building?
An MCP server is a small service that tells an AI client what it can do and then does it on request. The specification lists three things a server can offer: resources (context and data), prompts (templated messages and workflows) and tools (functions for the AI model to execute). A company's first server is almost always a handful of tools in front of an API it already has.
So the work is less about code than about choices: which operations agents may perform, how each is described, and who is allowed to call it.
What are the steps to build an MCP server for a company?
- Choose the jobs. Write down three to five things a customer or teammate would ask an agent to do with your product, such as find an order, create a ticket, or pull this month's usage.
- Turn each job into one tool. Give each tool a verb-first name, a description that says when to use it, and the smallest set of inputs. Return only the fields the model needs.
- Decide who may call it. For a remote server, require a key or an OAuth sign-in. Give each agent or teammate its own credential so you can revoke one without breaking the rest.
- Test it the way an agent sees it. List the tools, call each one with real inputs, and read the result as plain text. If you cannot tell what happened from the result, the model cannot either.
- Publish at a URL. Serve the server over Streamable HTTP so any client can reach it, then hand people a ready install snippet for their client.
- Watch the calls. Keep a log of every call with the tool, the status, the latency and the client. Fix the tools that fail and remove the ones nobody calls.
Three details from the specification are worth knowing before you start.
- Transport. The protocol defines two standard transports, stdio and Streamable HTTP. A server your customers reach over the internet uses Streamable HTTP, which exposes a single endpoint URL.
- Authorization. It is optional in the protocol and, for HTTP, based on OAuth 2.1. A server can also accept a static API key in an
Authorization: Bearerheader, which is what the install snippets for servers hosted on PlugMCP use. - Tool count. Tool definitions use the model's context. Anthropic describes a five-server setup whose 58 tools used about 55K tokens before the conversation started. Start small. More on this in the too-many-tools problem.
Should you write the server yourself or use a platform?
Both are reasonable. The right answer depends on who will maintain it.
| Write and run it yourself | Build and host on PlugMCP | |
|---|---|---|
| Who builds it | An engineer, with an MCP SDK | Anyone who can describe an API call |
| Where it lives | Your repo and your infrastructure | Your PlugMCP account; no repo |
| Running it | You deploy, scale and keep it up | The PlugMCP gateway runs it |
| Keys and call logs | You build them | Included |
| Logic between steps | Anything you can code | Workflow steps: branch, loop, filter, merge, call another tool |
| Best when | The server needs custom code or must run inside your network | The server is mostly API calls and you want it live quickly |
If your tools need arbitrary code or private network access, write it yourself. If they are calls to APIs and apps with some logic between them, a platform saves the hosting, the key handling and the logging.
What do those steps look like on PlugMCP?
The same six steps, with the infrastructure already there.
- Sign up and add a server. You choose one of three starts: host an MCP server that already exists, build one from an API or app, or combine servers you host into a toolkit behind one link.
- Describe each tool. For a single API call, fill in a form: method, address, inputs, and how to sign in to the API. When a tool needs more than one step, open it on the workflow canvas and connect steps such as call an API, if, switch, filter, loop, merge and call another MCP tool.
- Test before it is live. Workflow changes are saved as a draft that agents in production do not see. You can run the draft against test inputs before publishing, and roll back to an earlier published version if a change goes wrong.
- Publish. The server gets an endpoint on the PlugMCP gateway and an API key that starts with
pb_live_. The key is shown once. - Hand out the install. The console gives a ready snippet for Cursor, Claude Code, Claude Desktop, ChatGPT, VS Code, Windsurf, Codex, Gemini CLI and others, or a one-click link where the client supports one.
- Watch and fix. Every call is logged with its tool, status, latency and client. You can set an email alert for a rising error rate.
What if your server is already in the directory?
PlugMCP lists public MCP servers from the official registry, GitHub, npm, PyPI and other sources, so your server may have a page before you ever visit. Search for it in the directory and use Claim this server on its page. You prove you maintain it with your GitHub account, a DNS TXT record, or a file on your site. A claimed listing says who maintains it, and you can edit its details. If it is not listed, submit it.
A listing sends people to your own endpoint. If you also want PlugMCP keys and call logs in front of it, host it: see how hosting works.
What should you check before you announce it?
- Each tool name starts with a verb and each description says when to use the tool.
- No tool can delete or pay without that being obvious from its name and description.
- Every agent or teammate has its own key, and you know how to revoke one.
- You have called every tool with real inputs and read the result as the model would.
- There is an install snippet for the clients your customers use.
- You know where the call log is and who looks at it.
Background reading: why MCP exists, MCP vs A2A, and MCP Apps for tools that return a view.
Sources
Every figure and dated fact above comes from one of these pages. Follow the link to check it.
- Model Context Protocol specification, version 2025-06-18. modelcontextprotocol.io.
- MCP specification: Transports. modelcontextprotocol.io.
- MCP specification: Authorization. modelcontextprotocol.io.
- Introducing advanced tool use on the Claude Developer Platform. Anthropic Engineering, November 24, 2025.

