Cookies on softitland.com

We use essential storage to run this site. With your permission we would also like to use analytics cookies to see which pages help visitors, and marketing cookies to measure our ads. You can change your choice at any time with “Cookie settings” in the footer. Privacy Policy

AI & automation

MCP explained: connecting AI assistants to your CRM, files and databases

What MCP (Model Context Protocol) is, how it connects AI assistants to your CRM, files and databases, sign-in and risks, and when a plain API is simpler.

Cover image for the article MCP explained: connecting AI assistants to your CRM, files and databases
Key takeaways
  • MCP is an open standard: one server per system lets any compatible AI app, such as Claude or ChatGPT, use that system under your permissions.
  • Start read-only, with per-user sign-in and the smallest scopes, and add write tools one at a time behind approval.
  • Treat outside MCP servers as software with access to your data; tool descriptions from untrusted servers can carry hidden instructions.
  • For one fixed automation, a workflow that calls the API directly is usually simpler than an MCP server.

What is MCP, and when should you use it?

MCP, the Model Context Protocol, is an open standard for connecting AI applications to outside systems. You build one MCP server for a system such as your CRM, database or file store, and any compatible AI app, such as Claude or ChatGPT, can use it with the permissions you allow. Use it when several people or AI tools need the same connection, or when your team wants company data inside the AI apps they already use. For one fixed automation, a direct API call is often simpler.

The official introduction describes MCP as an open-source standard for connecting AI applications to external systems. Anthropic released it in 2024, and since December 2025 it has sat in the Agentic AI Foundation, which the MCP blog describes as a directed fund under the Linux Foundation, co-founded by Anthropic, Block and OpenAI.

How does MCP work?

There are three roles, according to the architecture guide:

  • Host: the AI application the person uses, such as Claude Desktop or an internal assistant.
  • Client: the part of the host that holds a connection to one server.
  • Server: a program that gives the AI context and actions from one system, such as your CRM.

A server can offer three kinds of things: tools the model can call (search contacts, create a note), resources it can read (a file, a record) and prompts, which are reusable templates. Servers run either locally over standard input and output, or remotely over Streamable HTTP, as set out in the transports spec. The current protocol version is 2026-07-28, and the spec still changes, so plan for updates.

Which AI apps can use an MCP server?

  • Claude: custom connectors using remote MCP are available on Claude, Cowork and Claude Desktop, per Claude's help centre. Claude connects to your server from Anthropic's cloud, not from your own computer, so a remote server must be reachable over the internet.
  • OpenAI: the OpenAI connectors and MCP guide lets the Responses API call remote MCP servers, and by default OpenAI requests your approval before any data is shared with one.
  • Your own software: an internal assistant or agent you build can act as an MCP host too, so the same server serves staff tools and customer-facing ones.

That reuse is the main business case. Build the CRM connection once, with its permissions and logs, and every approved AI app uses the same controlled door instead of each tool getting its own copy of your API keys.

What should the first connection expose?

Design tools around the jobs people do, not around every API endpoint. A few patterns we use:

  • CRM: search contacts and companies, read a deal with its recent activity, draft a note. Creating or changing records comes later, behind an approval step.
  • Database: read-only queries against views made for the purpose, with fixed parameters, never free-form SQL written by the model.
  • Files: resources limited to the folders a role may see, with the source path returned so people can check it. For large document sets, compare this with retrieval in our guide to RAG, fine-tuning or a long prompt.

Start read-only. Most of the value, such as finding a customer's history in seconds or summarising open deals, needs no write access at all.

How do sign-in and permissions work?

The authorization spec makes authorization optional in the protocol, defines an OAuth 2.1-based flow for HTTP servers, and says local servers should take credentials from their environment instead. For anything that touches business data, treat sign-in as required.

  • Per-user sign-in: each person connects with their own account, so the assistant sees only what that person can see in the CRM or file store.
  • Token rules: the spec says servers must check that each access token was issued for them, and must not accept or pass on any other tokens.
  • Smallest scopes: clients should request only the scopes they need, and the security best practices recommend a progressive, least-privilege scope model.

What are the security risks?

The biggest risk is an AI model being steered by text it reads. Claude's help centre warns that malicious MCP servers may include hidden instructions that try to make Claude perform unintended actions, and the tools spec says clients must treat tool annotations as untrusted unless they come from trusted servers.

The specification says hosts must obtain explicit user consent before invoking any tool, but it also says the protocol cannot enforce these principles itself. That job falls to the people building the host and the server. In practice:

  • Connect only servers from publishers you trust, and review the permissions each one asks for.
  • Keep a person approving any tool that writes, sends or deletes, as covered in our guide to guardrails for AI agents.
  • Log every tool call with the user, the inputs and the result.
  • Keep customer data rules in force; see keeping customer data safe with AI APIs.

When is a plain API integration simpler?

MCP is a connection standard, not a requirement. A single automation that always does the same thing, such as reading a form, summarising it and creating a CRM note, is usually simpler as a workflow that calls the AI model and the CRM API directly, built in n8n, Make or a small service. There is one connection to maintain and no protocol layer to update.

MCP earns its place when the same system must be reachable from several AI apps, when people should pick tools in conversation instead of following a fixed flow, or when your team wants company data inside Claude or ChatGPT under your own permissions.

How should you scope a first MCP project?

  • Pick one system and one team, such as sales reading the CRM.
  • List five to ten questions that team asks every week, and build read-only tools that answer them.
  • Add per-user sign-in with the smallest scopes, plus logging, before anyone uses it.
  • Test with real questions and with hostile text in records, then add write tools one at a time behind approval.

Our AI agents and integrations team designs and hosts MCP servers and the assistants that use them, our CRM and lead systems team prepares HubSpot or GoHighLevel data and permissions, and our custom software team builds servers for in-house databases and portals.

Frequently asked questions

Is MCP only for Claude?

No. Anthropic released MCP, but it is an open standard now held in the Agentic AI Foundation under the Linux Foundation, and other apps support it, including OpenAI's API and ChatGPT. A server built to the spec can be used by any compatible client, which is the main reason to build one.

Do we need to build our own MCP server?

Not always. Some software vendors publish their own MCP servers, so check whether yours does and what permissions it requests. Build your own when the system is in-house, when the vendor's server exposes too much or too little, or when you need your own logging and approval rules.

Can the AI change records in our CRM through MCP?

Only if the server offers write tools and the signed-in user's permissions allow the change. The specification says hosts must get explicit user consent before invoking tools. We start with read-only tools and add writes one at a time, each behind a person's approval.

Is it safe to connect a third-party MCP server?

Treat it like installing software that can read your data. Malicious servers can hide instructions in what they return, so check the publisher, read the permissions it asks for, test it with a limited account, and keep approval on any action that writes, sends or deletes.

What does an MCP server need to run in production?

Hosting over HTTPS that the AI app can reach, sign-in with OAuth, logs of every tool call, monitoring and someone responsible for updates. The protocol is still evolving, with the current version dated 2026-07-28, so budget time to keep the server in line with new releases.

Last updated

Start a project

Have a project in mind? Let’s build it.

Tell us what you need. You get a clear plan, a fixed quote and a timeline before any work starts.

Prefer email? info@softitland.com