Quick answer: An MCP server is a small connector that lets an AI assistant such as Claude or ChatGPT use one of your work tools, like Gmail, HubSpot or your ATS. It gives the assistant a fixed list of actions (search contacts, read an email, create a draft) and nothing more. MCP stands for Model Context Protocol, an open standard Anthropic introduced in November 2024.
What it means for recruiters: Instead of copying a job description out of your ATS and pasting it into a chat, the assistant reads it directly. The server decides what the assistant can see and do, and your login decides which records it can reach.
This guide is part of our MCP servers for recruiters series. It explains the idea without code, so you can judge any connector a vendor or colleague puts in front of you.
What is an MCP server, in plain English?
An MCP server is a translator that sits between an AI assistant and a piece of software. The assistant asks for something in ordinary language, and the server turns that into an action the software understands, such as “find contacts at this company” or “list my events on Tuesday.”
Think of it as a menu. Each MCP server publishes a short list of things it can do, called tools. The assistant reads the menu, picks the tool that fits your request, and the server carries it out and hands back the result.
That menu is the important part. A server only does what its tools allow. If a Gmail server has no send tool, the assistant cannot send an email through it, however you phrase the request.
What does MCP stand for, and who created it?
MCP stands for Model Context Protocol. “Model” is the AI model, “context” is the information it needs to do a job, and “protocol” is the agreed set of rules for passing that information back and forth.
Anthropic, the company behind Claude, introduced MCP in November 2024 as an open standard. Open means any company can build a server or a client without asking permission, and the rules are published on the MCP spec site.
That openness is why the same idea now shows up across assistants. ChatGPT, Claude, GitHub Copilot CLI and coding tools like Cursor can all talk to MCP servers, although each one names and packages them differently.
How does an MCP server work?
Every MCP setup has three parts, plus your sign in. Once you know them, every vendor announcement makes more sense.
| Part | What it is | Recruiting example |
|---|---|---|
| Client | The AI assistant you type into | Claude, ChatGPT, Cursor |
| Server | The connector run by the tool’s vendor or a third party | HubSpot’s MCP server, Apollo’s MCP server |
| Tools | The individual actions the server offers | Search contacts, create a deal, draft an email |
| Sign in | Your own login, usually through OAuth | You approve access once with your HubSpot or Google account |
Here is what happens when you ask Claude to pull last week’s new contacts from your CRM.
1. You ask in plain language
You type “Show me contacts added to HubSpot since Monday at software companies.” No special syntax is needed.
2. The assistant picks a tool
The assistant checks the tools the connected server offers and chooses the one that searches contacts. It fills in the filters from your request.
3. The server runs it under your login
The server sends the request to the software using your own account. You only see records your account can already see. Salesforce, for example, says each call to its hosted servers runs with the signed in user’s permissions.
4. The result comes back to the chat
The server returns the records and the assistant summarizes them, builds a table, or uses them in the next step, such as drafting a follow up note.
What is the difference between an MCP server and an API?
An API is a doorway built for developers. Someone has to write code to call it, handle errors and decide what to do with the data. An MCP server is a doorway built for AI assistants. The assistant reads the tool list and calls it on your behalf, so you do not write any code.
Most MCP servers sit on top of an existing API. The vendor picks the API actions it is comfortable exposing, describes them in plain words, and wraps them as tools. That is why a tool can have a full API but a smaller MCP server.
| Question | API | MCP server |
|---|---|---|
| Who uses it? | Developers writing code | AI assistants acting for you |
| Do you need to code? | Yes | No, you connect it in settings |
| How do you sign in? | API keys or tokens set up by a developer | Usually your normal login through OAuth |
| What can it do? | Everything the vendor documents | Only the tools the vendor chose to expose |
| Best for | Building integrations and reports | Asking questions and doing tasks from a chat |
For recruiters, the practical takeaway is simple. “Has an API” does not mean you can use it in Claude or ChatGPT. Look for “MCP server” or “connector” in the vendor’s docs. Our live MCP tracker counts only tools whose vendor runs its own server.
What is the difference between a local and a remote MCP server?
A local MCP server runs on your own computer. It is usually installed by a developer and is common in coding tools. A remote MCP server runs on the vendor’s own systems at a web address, and you connect it by pasting that address or clicking a button in your assistant’s settings.
The vendor servers recruiters are likely to meet are remote, which suits a recruiting desk where nobody wants to install software to try a connector. HubSpot, for example, runs a remote HubSpot server that became generally available on April 13, 2026.
| Local server | Remote server | |
|---|---|---|
| Where it runs | Your laptop | The vendor’s systems |
| Setup | Install and configure files | Paste a URL or click connect, then sign in |
| Who maintains it | You or whoever wrote it | The vendor |
| Typical user | Developers | Everyone else, including recruiters |
Is a connector the same thing as an MCP server?
Mostly, yes. Claude calls them connectors, ChatGPT calls them apps, and both can run on MCP behind the scenes. The words differ but the idea is the same: a link from your assistant to one of your tools.
One distinction is worth knowing. Some connectors are built by the AI company, and some are built by the tool’s vendor. Claude’s own Gmail connector, made by Anthropic, can send email and asks before each send by default. Google’s own Gmail MCP server is a developer preview with no send tool at all. Same inbox, different menus. Our Gmail and Outlook guide covers both.
What can an MCP server actually do for a recruiter?
It depends entirely on the tools a vendor exposes. Here is what a few servers that recruiters use every day offer today, as documented by the vendors.
| Tool | What its server can do | Worth knowing |
|---|---|---|
| Gmail (Google) | Search and read threads, create drafts, manage labels | Developer preview; no send tool |
| HubSpot | Read and update contacts, companies, deals and more | No delete in Claude |
| Apollo | Search people and companies, enrich, manage sequences | Can send emails and spend credits |
| Microsoft 365 (Claude) | Read Outlook mail, calendar, Teams, SharePoint, OneDrive | Read only by default |
| Workday | Agent Ready Tools over MCP | Early Availability; recruiting not mentioned |
In day to day terms, that covers jobs like these: summarize every email from a hiring manager this week, list open deals with no activity in 14 days, find VPs of engineering at a target account, or draft a candidate update you review before it goes out. Our guides on sourcing with Claude and finding clients with Claude show full workflows.
Is it safe to connect an MCP server to candidate data?
It can be, if you set it up with care. The risk is not the protocol itself. The risk is giving an assistant write or send access to systems full of personal data and then not checking what it does.
Four habits cover most of the risk.
1. Keep write and send actions on approval
Vendors recommend this themselves. HubSpot suggests putting write tools on “Needs Approval”, and Apollo recommends setting credit using actions to “Approval required”. That way nothing is sent, changed or charged without a click from you.
2. Let your admin approve access
On team plans, an admin usually has to switch a connector on. Microsoft 365, for example, needs an Entra Global Administrator to grant tenant consent once. That step exists so the business decides what leaves its systems.
3. Start read only
Connect in read only mode first and learn how the assistant behaves before you allow writes. Claude’s Microsoft 365 connector starts this way by default.
4. Treat candidate data with care
Only connect the accounts you need, avoid pasting sensitive candidate details into personal plans, and check your local rules before letting AI rank or screen people. Our AI hiring law checker is a good place to start.
How do you connect your first MCP server?
The steps are similar in every assistant. In Claude, look under connectors in settings. In ChatGPT, look under apps. If a vendor is listed, click connect and sign in. If it is not listed, you can usually add it as a custom connector by pasting the server URL the vendor publishes.
A low risk way to try it is our own Work Insiders MCP server. It is free, read only, and covers our directory of recruiting tools, so you can practice without touching candidate or client data. It works in Claude, ChatGPT and Cursor.
Claude prompt: check a connector before you turn it on
I am a recruiter thinking about connecting [TOOL NAME] to you through its MCP server. List every tool the connector gives you, then sort them into three groups: read only, creates or changes records, and sends messages or spends credits. For each write or send tool, tell me what could go wrong if I approved it by mistake, and which ones I should keep on approval. Do not run any of the tools.
For a list of recruiting tools that already ship a server, with the URLs to paste, see our Claude connectors list. For the everyday tools most recruiters use, start with the MCP servers hub.
Quick answers
What does MCP server mean?
It is a connector that lets an AI assistant use a specific piece of software through a fixed list of actions. MCP stands for Model Context Protocol.
Do I need to code to use an MCP server?
No. Remote servers connect from your assistant’s settings with your normal login. Coding is only needed to build a server or run a local one.
Is an MCP server the same as an API?
No. An API is for developers writing code. An MCP server is for AI assistants, and it usually exposes only part of what the API can do.
Can an MCP server see everything in my account?
It can only use the tools it offers, and those run under your login. You will not see records your account cannot already see.
Does every recruiting tool have an MCP server?
No. Many do not yet, and some have only third party servers. Our live tracker keeps the count.
Which assistants support MCP?
Claude, ChatGPT, GitHub Copilot CLI and coding apps such as Cursor all work with MCP servers, though setup and plan rules differ by vendor.
The bottom line
An MCP server is how an AI assistant gets hands inside one of your tools. It can only do what its tools list allows, it runs under your login, and the safe way to start is read only with sends and writes on approval. Pick one tool you live in every day, connect it, and see how much copying and pasting disappears. The MCP servers for recruiters hub shows which everyday tools are ready.
Sources: Anthropic MCP announcement, MCP spec site, Google Gmail MCP, Claude Google connectors, Claude Microsoft 365, HubSpot MCP GA, HubSpot Claude connector, Apollo MCP docs, Salesforce hosted servers, Workday agentic blog
Free newsletter
Get the Work Insiders weekly
One email a week: the recruiting and HR tech news that mattered, the tools we just added to the directory, and one practical guide. No spam, and you can unsubscribe in one click.