An MCP server is a program that gives an AI assistant such as Claude or ChatGPT access to the data and actions of another system through the Model Context Protocol (MCP), an open standard. For a sales team, that other system can be a company database, a contact enrichment service, a CRM or a campaign tool, so a rep asks for accounts, contacts or a sequence in plain language instead of copying records between tabs. Connecting one takes three things: a remote server reachable over the internet, an assistant that supports MCP, and a sign-in step, usually OAuth.
What is an MCP server?
An MCP server is a program that exposes data and actions to an AI assistant such as Claude or ChatGPT through the Model Context Protocol, an open standard, so the assistant can query that data and act on it without anyone copying and pasting.
Anthropic open-sourced the Model Context Protocol on 25 November 2024 and described it as "a new standard for connecting AI assistants to the systems where data lives." The current version of the MCP specification published on modelcontextprotocol.io is dated 28 July 2026.
The three parts of MCP: host, client and server
The Model Context Protocol splits every connection into three parts, according to the architecture page of the official MCP documentation:
- MCP host: the AI application the person works in, such as Claude or ChatGPT, which coordinates one or more MCP clients.
- MCP client: the component inside the host that keeps a connection open to one MCP server and fetches context from it.
- MCP server: the program that provides that context, for example a prospecting database or a CRM.
A local MCP server runs on the user's own computer and usually serves a single client. A remote MCP server runs on the internet over the Streamable HTTP transport and usually serves many clients.
What an MCP server exposes: tools, resources and prompts
An MCP server exposes three kinds of building blocks, which the MCP documentation calls primitives:
- Tools: functions the assistant can run, such as "search companies" or "create contact".
- Resources: data the assistant can read for context, such as a company record or a saved list.
- Prompts: reusable templates that structure a task, such as a qualification checklist.
For sales work, the tools of an MCP server matter most, because they are what lets the assistant act. A read-only tool looks something up; a write tool changes a record, enrols a contact or starts a paid search.
What sales teams use an MCP server for
Sales teams use an MCP server to run the steps of prospecting from inside Claude or ChatGPT: finding target accounts, finding the right people in them, enriching those people with an email and a phone number, launching outreach and checking what the CRM already holds. Each step is one or more tools on a server.
- Account search: an account search tool builds a list of companies from filters such as industry, location and company size.
- Contact search: a contact search tool finds people by job title, seniority or function inside the accounts on that list.
- Enrichment: an enrichment tool adds a work email and a phone number to each contact, the step covered in our guide on how to find someone's phone number.
- Campaigns: a campaign tool enrols the chosen contacts in a LinkedIn or email sequence, the core of outbound sales.
- CRM lookups: a CRM tool checks whether an account is already a customer or an open deal before anyone writes to it.
A typical request from an SDR selling to logistics companies looks like this:
Find logistics companies in Amsterdam with 51 to 200 employees, then the Head of Operations at each one, and get their work email.
The assistant splits that request into a company search, a contact search and an enrichment call. The process underneath is the one described in our guide to sales prospecting: an MCP server changes where the rep runs it, not what good prospecting looks like.
MCP server vs API, integration and browser extension
An MCP server differs from an API, a classic integration and a browser extension in who uses it and who decides what runs. An API is built for developers' code, an integration moves data on a fixed schedule, and an extension works inside one web page. An MCP server is built for an AI assistant, which picks the tool and fills in the details from a request in plain language.
| Option | Who uses it | What it takes to connect | Who decides what runs |
|---|---|---|---|
| MCP server | An AI assistant acting for a person | Add the server URL to the assistant and sign in, usually with OAuth | The assistant, within the tools the server exposes and the permissions the person grants |
| API | A developer's code | Write and host code against the API's endpoints and keys | The developer, in code |
| Integration (sync) | Two applications, with no person involved | Map fields and set a sync schedule | Rules fixed at setup |
| Browser extension | A person, on the page that is open | Install the extension in the browser | The person, click by click |
Many MCP servers are a thin layer over an existing API: the API still does the work, and the MCP server describes it in a form the assistant understands. That is why connecting a sales tool through its MCP server needs no code on the sales team's side.
How to connect a company database to Claude
There are three ways to connect a company database to Claude, from least to most engineering work: add an MCP server that already holds company and contact data, put an MCP server in front of your own database, or write your own code on Anthropic's API. Claude reaches remote MCP servers from Anthropic's cloud, so the server must be reachable over the public internet, according to the Claude Help Center.
- An MCP server that already has the data. A prospecting MCP server brings its own company search and contact data, so connecting it means adding it to Claude as a custom connector and signing in. Custom connectors using remote MCP are available on Claude's Free, Pro, Max, Team and Enterprise plans, and Free users are limited to one custom connector, according to the Claude Help Center.
- An MCP server over your own database. When your accounts already live in your own Postgres database or warehouse, an MCP server in front of it lets Claude query them. The MCP project published a PostgreSQL reference server with read-only access and schema inspection; it is now archived, and the MCP Registry lists published servers to choose from instead.
- Your own code on Anthropic's API. A developer can call Claude through the API and define the database queries as tools.
On Claude Team plans, only Owners and Primary Owners can add a custom connector for the organisation; each member then connects it with their own account.
How to connect the same data to ChatGPT
ChatGPT connects to the same remote MCP servers, packaged as plugins: OpenAI's product documentation states that ChatGPT on the web uses remote MCP tools through plugins.
According to OpenAI's developer documentation, a custom server is added from the Plugins page in ChatGPT with "Add custom MCP server": you give it a name, enter the server URL, configure authentication and review the tools ChatGPT discovers. The server needs a public HTTPS endpoint, or OpenAI's Secure MCP Tunnel when it sits on a private network.
What to check for security and cost before connecting a sales MCP server
Before connecting a sales MCP server, check three things: how it signs you in, what it is allowed to change, and what stops it from spending more than planned.
- Sign-in: the MCP specification bases authorization for remote servers on OAuth 2.1, so the assistant receives a scoped token and never sees your password. Authorization is optional in the protocol, so confirm that the server you connect uses it.
- Write permissions: MCP tools can read, create, modify or delete records. The Claude Help Center recommends reviewing the permissions a server requests, disabling tools a conversation does not need and choosing "Allow always" only for tools you trust to run unsupervised.
- Spend controls: a server that runs paid searches or enrichment should show the cost before it starts and cap what one organisation can spend. In HappySales, for example, every paid step is valued before it runs, and a daily spending cap per organisation, together with rate limits, keeps an agent from spending more than planned.
The Claude Help Center adds that a malicious server can carry hidden instructions, which is why the safe default is to connect only servers built by organisations you trust.
A sales MCP server in practice: HappySales as an example
HappySales is a prospecting MCP server whose flow has four steps, all run from inside Claude or ChatGPT: a company search, a contact search in the companies found, email and phone enrichment, and a campaign. The rep asks for each step in plain language and the assistant calls the matching HappySales tools; the same flow is available in the HappySales web app.
- Company search in HappySales: the search runs on LinkedIn's company search, with single keywords, as many as 20 locations, as many as 20 industries from LinkedIn's taxonomy of 434 and LinkedIn's headcount bands. A single search tops out at 1,000 companies, a limit set by LinkedIn rather than by HappySales; a bigger market is covered by splitting it by location or headcount.
- Contact search in HappySales: a contact search returns as many as 100 people from each company and 10,000 in total, filtered by title, function, seniority or a recent move to a new job, and stores them in contact lists.
- Enrichment in HappySales: each run enriches as many as 100 contacts through a waterfall over several data providers; a contact is billed only when an email or phone is actually found, and every email comes back labelled deliverable, catch-all or undeliverable.
- Campaigns in HappySales: the chosen contacts go into one outreach sequence that combines LinkedIn and email.
The connection works the way this guide describes for any remote server: the team adds HappySales to Claude or ChatGPT as a connector, and sign-in uses OAuth 2.1. For access, contact HappySales.
HappySales is not a CRM: it covers the work that happens before a deal exists, with no pipeline, quote or forecast management, so a team that wants the assistant to update deals connects its CRM's own server alongside.