Assistants
MCP server, or an API?
Nexora ships MCP servers in production: the assistant inside Extrack has 29 actions behind it, each one a thing it can carry out rather than describe, live at myextrack.com. The question underneath is not which protocol is better. It is whether the thing calling your system is a program that already knows what it wants, or a model deciding in the moment.
What is the actual difference?
An API is called by software that already knows which endpoint it needs. An MCP server is read by a model that has to choose, from a list, without being told. That inverts what the interface is for: an API is a contract between two programs, and an MCP tool definition is an explanation aimed at a reader who will decide based on how it is written.
Do I need an MCP server, or is my API enough?
If the calls are made by your own code on a path you control, the API is enough and an MCP server adds a layer for nobody. You need one when you want an assistant working against your real data and choosing which operation fits — which is what Extrack's 29 actions are. Wrapping an existing API is usually where it starts.
Why do tool descriptions matter so much?
Because the description is the entire basis on which a model picks. Two tools that read alike get confused for each other, and the failure looks like a bad model rather than a bad schema. The free MCP tool-schema auditor on this site exists because that was the problem worth checking first.
What does the tool list cost before anyone speaks?
Every definition travels with every request. Anthropic's own tool-use documentation is explicit that pricing counts the tools parameter, and that the extra tokens come from names, descriptions and schemas. A large tool list is a standing cost on every message, which is why the auditor reports it.
How long does one take to build?
An AI assistant is a three-to-four-week build. Extrack went from idea to first paying customers in 15 days with its assistant included, so the band is a real one rather than a sales number.
Questions we get asked
- Can an MCP server wrap an API we already have?
- Usually, and that is the normal starting point. The work is in the tool definitions rather than the plumbing: the endpoints already exist, and what has to be written is the explanation a model reads to choose between them.
- Does an MCP server replace our API?
- No. It sits in front of one. Your own software keeps calling the API directly, because it already knows which call it wants.
How Extrack's 29 actions were built