Agents and MCP · 6 MIN
Does your product need an MCP server? A guide for SaaS and product teams
When a SaaS or product company should build an MCP server so AI assistants can use its product, what it exposes, how authorization works, what it costs to run, and how to decide.
Your product needs an MCP server when your customers already use AI assistants such as Claude, ChatGPT or Cursor for work your product is part of, and they would rather ask the assistant than open your app. An MCP server lets those assistants read your product's data and take actions in it, under the customer's own login. Companies like GitHub and Atlassian already offer remote MCP servers. If your product has an API that customers use for daily work, the MCP server builds on top of it rather than replacing anything.
- MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems, often described as a USB-C port for AI.
- An MCP server can offer three things: tools (actions), resources (data for context) and prompts (ready-made templates).
- Authorization in MCP is built on OAuth 2.1, so customers can connect with their own account and permissions.
- GitHub hosts a remote MCP server, and Atlassian launched one for Jira and Confluence in May 2025.
- Build one when customers already ask AI tools about work your product holds. Skip it when nobody would use it daily.
- Nactore is a technology and growth company that designs and builds MCP servers for products, then helps customers find them.
What is an MCP server, in plain terms?
The MCP introduction describes MCP as an open-source standard for connecting AI applications to external systems, and compares it to a USB-C port: one standard plug instead of a custom cable for every device.
An MCP server is your product's side of that plug. Once it exists, any AI application that supports MCP can connect to your product without you building a separate integration for each one. For a longer background, read what MCP is.
What can an MCP server expose?
The MCP specification defines three building blocks a server provides:
| Building block | What it is | Example for a project management product |
|---|---|---|
| Tools | Functions the model can call to take actions or fetch information | "Create a task", "move a task to Done" |
| Resources | Data the application can give the model as context | The current sprint's tasks, a project's description |
| Prompts | Templates the user picks, such as slash commands | "Summarize this week's progress" |
Most product teams start with a handful of tools for the actions customers do most, plus read access to the data those actions need.
Who already offers one?
- GitHub publishes the GitHub MCP Server, including a remote version hosted by GitHub that works with hosts such as VS Code, Claude Desktop and Cursor.
- Atlassian announced a remote MCP server on 1 May 2025, starting with Anthropic as its first official partner and hosted on Cloudflare, so Jira and Confluence Cloud customers can summarize, create and update work items and pages from Claude.
When the tools your customers already use offer MCP servers, it is fair to expect them to ask the same of yours.
When is it worth building?
| Signal | Build now | Wait |
|---|---|---|
| Customers ask about AI integrations in sales calls or support | Yes | |
| Your product has a stable, documented API | Yes | |
| Users do repeated actions that are easy to describe in words | Yes | |
| Your users are developers or technical teams using AI coding tools | Yes | |
| Your API is changing every month | Yes, stabilize it first | |
| Your product is used once a quarter | Yes, low daily value | |
| Data is so sensitive that customers would not connect an assistant | Yes, until controls exist |
How does authorization work?
The MCP authorization spec makes authorization optional for MCP implementations, but where it is used, it is built on OAuth 2.1: the AI application acts as an OAuth client making requests on behalf of the user. For a SaaS product this is the right model. Each customer signs in with their own account, the assistant gets the same permissions that user already has, and your existing access rules keep working.
Do not ship a server that uses one shared API key for everyone. Our guide to MCP security risks covers what else to check, including prompt injection through data the assistant reads.
What does building one involve?
- Pick the first five to ten tools. Start from the actions customers do most, not from every API endpoint.
- Design tools for a model, not a developer. Clear names, short descriptions, small inputs and helpful errors. See tool-calling design for agents.
- Wire up OAuth so each user connects their own account.
- Decide hosting. A remote server you host gives every customer the same version and lets you see usage. A local server runs on the customer's machine.
- Test with real assistants. Run real tasks in two or three MCP hosts and fix what confuses the model.
- Write the setup page. One page with copy-paste setup for each popular host.
- Measure use. Calls per tool, errors and active customers tell you what to build next.
Our field notes on building an MCP server cover the engineering details we learned the hard way.
How do customers find out about it?
An MCP server nobody knows about does nothing. Add it to your docs, your integrations page and your changelog, and list it in the directories developers check. Write the setup page so AI assistants can read it too, because a customer may ask an assistant "does [product] have an MCP server?" before they search.
Frequently asked questions
Is an MCP server the same as an API?
No. It sits on top of your API and describes your product in a way AI applications understand, so they can decide which action to call and with what inputs.
Does MCP only work with Claude?
No. It is an open protocol supported across many clients. GitHub's server, for example, lists VS Code, Claude Desktop, Cursor and Windsurf among compatible hosts.
How much does it cost to run?
A remote server is a web service in front of your existing API, so it adds hosting like any other small service. Most of the effort is in designing good tools and wiring up sign-in.
Can the assistant do anything a user can do?
Only what your tools expose, and only with the signed-in user's permissions. Start with a narrow set and add more as you see how customers use it.
Make your product usable by AI
When customers do their work through assistants, the products that plug into those assistants stay part of the routine. Nactore designs and builds MCP servers as part of AI-native engineering, and makes sure customers and AI tools can find them.
Want to scope an MCP server for your product? Book a free 30-minute call.
Want to apply this to your business?
Book a free 30-minute call. We will tell you what we would do first.