---
title: "MCP in recruitment: definition, uses and limits"
description: "MCP connects Claude or ChatGPT to your ATS to search, create and follow up with candidates in plain language. Definition, uses and limits."
url: https://marvins.ai/en/blog/mcp-recruitment-ai-ats
language: en
published: 2026-04-14
updated: 2026-10-23
author: "Christophe Hébert"
---

# MCP in recruitment: definition, uses and limits

MCP connects Claude or ChatGPT to your ATS to search, create and follow up with candidates in plain language. Definition, uses and limits.

**MCP (Model Context Protocol)** is the standard that connects a conversational AI, Claude or ChatGPT, to business tools such as your ATS. A recruiter asks for something in plain language, the AI queries your talent pool, your open roles and your pipeline through your recruitment software's MCP server, then carries out the requested action. Marvin exposes this server today, for every client.

**_Article updated on 23 October 2026_**

## What to remember

- **MCP is a protocol, not a product.** Anthropic published it as an open standard in late 2024; any software publisher can implement it.
- **An MCP server is the door a publisher opens onto its tool.** Marvin exposes its own, with 60 recruiter tools reachable from Claude, ChatGPT or another compatible assistant.
- **Recruiters keep their rights.** The AI sees what the connected Marvin account sees, nothing more: tenant isolation applies to every call, exactly as it does inside Marvin.
- **Nothing goes out without confirmation.** Sending a message, deleting a record or merging profiles goes through a preview before it runs.
- **MCP covers the whole recruiter job**, not only sourcing: CRM, ATS, outreach, assignments, messaging, reporting.

## What is MCP (Model Context Protocol)?

The Model Context Protocol is an open standard published by Anthropic in late 2024. It describes how a generative AI calls the functions of a business tool: look up a record, create a profile, trigger an action, without custom development or data extraction outside the tool.

Before MCP, connecting an AI to a piece of software meant building a dedicated integration for each tool-assistant pair. MCP replaces these one-off integrations with a single contract: a publisher describes the available actions once, and any compatible client (Claude, ChatGPT, Cursor...) uses them the same way.

The protocol does not keep any data in permanent storage. Each exchange is a request opened by the recruiter, handled in real time, closed at the end of the conversation. For a recruiter, the practical gain is direct: fewer tabs to open, fewer manual exports, and no copy-pasting between the ATS, LinkedIn and a reporting spreadsheet to answer a simple question.

## What is an MCP server for an ATS?

An MCP server is a publisher's implementation of the protocol. For an ATS, it exposes the actions a recruiter already performs in the tool: search for a candidate, read a record, create a process, send a message, update a status. The AI connects to the server with the recruiter's own account, through OAuth, then turns a plain-language request into the right calls.

Read actions, such as searching or reading a record, and write actions, such as creating, updating or sending something, are both covered by the same server. A write action still goes through the same approval step a recruiter would apply manually, which keeps the human decision in the loop.

Marvin publishes its MCP server at a single address, `https://mcp.marvins.ai/mcp`. The OAuth connection determines each user's account, agency and rights; the address itself grants access to nothing.

## What can a recruiter do with an MCP server?

The concrete uses span the whole job, not only candidate search.

| Use                      | Example request                                                            | What the AI does                                                                         |
| ------------------------ | -------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------- |
| Search the talent pool   | "Find accountants available in January, based in Lyon."                    | Queries your candidate database with the given filters and returns a ranked list.        |
| Create a record          | "Here's a CV, create the candidate record with role, skills and location." | Reads the document, extracts the relevant fields and creates the record in your ATS.     |
| Check on a pipeline      | "Where do candidates stand on the Payroll Manager role?"                   | Reads the pipeline and the message history, then summarises the state of the process.    |
| Personalised follow-up   | "Follow up with candidates who haven't replied in 5 days on this process." | Drafts one message per candidate and submits it for your approval before anything sends. |
| Conversational reporting | "How many placements this quarter, by recruiter?"                          | Queries your data and answers in text, without opening a dashboard.                      |
| Clean up the taxonomy    | "List the duplicate skills in our taxonomy."                               | Finds duplicates, counts their impact and proposes a merge submitted for your approval.  |

In every case, the AI relies on the same data and the same rights as the connected Marvin account. It doesn't guess anything: it reads what the recruiter can read, and acts within what the recruiter can do.

## What are the limits and the security model?

An MCP server touches sensitive data: candidates, salaries, client exchanges. Several guarantees frame how it can be used.

- **Every recruiter connects with their own account**, through OAuth. No shared key, no password handed to the AI.
- **Marvin's visibility rights apply to every call.** The AI only sees what the connected account sees, and isolation between agencies stays strict.
- **Sensitive actions go through a preview.** Sending a message, deleting a record, merging profiles: the AI shows what it's about to do, and the recruiter confirms before it runs.
- **Data stays hosted in France**, at Marvin. Only the data you ask for travels to the AI service you use, for the length of the conversation; Marvin doesn't replicate its database to a third party.
- **Every action stays traceable.** What the AI does runs through the recruiter's own account, with the recruiter's own rights, and is logged like any other action taken inside Marvin.
- **One real limit remains**: MCP depends on the quality of the AI client in use. A poorly guided assistant can misread an ambiguous request. Reviewing a message before approving it is still necessary, as with any text an AI drafts.

## How do you use it with Marvin?

Marvin publishes [Marvin's MCP server to connect your ATS to Claude or ChatGPT](https://marvins.ai/en/mcp) from a single address. Setup takes two steps, in your account settings, with no code to write.

Recruiters who'd rather stay inside the Marvin interface than go through an external AI get the same outcome with [Marvin AI's Copilot](https://marvins.ai/en/ai), directly in the platform.

**[Book a demo](https://marvins.ai/en/demo)** to see MCP on your own data.

## FAQ

**Does MCP replace my ATS?**

No. MCP is a connector: it gives your AI access to the actions your ATS already supports, it doesn't store or replace your data. Marvin stays the system where your records, your pipeline and your history live; MCP simply opens a door onto it.

**What is the difference between an AI agent and an MCP server?**

An MCP server describes the actions an AI can take inside a tool: search for a candidate, create a record, send a message. An agent is the program that decides, on its own, when and how to use them. Claude or ChatGPT play that agent role when they connect to the Marvin MCP server.

**Does MCP work with any AI?**

With any AI that supports the standard: Claude, ChatGPT, and developer tools such as Cursor. The address of Marvin's MCP server is the same for every compatible client; only the OAuth connection changes depending on the account used.

**Is MCP only useful for sourcing?**

No. Sourcing is only one of the possible uses. Marvin's MCP server also covers CRM, ATS, outreach, assignments and timesheets, messaging and reporting: the actions a recruiter already performs in Marvin.

## FAQ

### Does MCP replace my ATS?

No. MCP is a connector: it gives your AI access to the actions your ATS already supports, it doesn't store or replace your data. Marvin stays the system where your records, your pipeline and your history live; MCP simply opens a door onto it.

### What is the difference between an AI agent and an MCP server?

An MCP server describes the actions an AI can take inside a tool: search for a candidate, create a record, send a message. An agent is the program that decides, on its own, when and how to use them. Claude or ChatGPT play that agent role when they connect to the Marvin MCP server.

### Does MCP work with any AI?

With any AI that supports the standard: Claude, ChatGPT, and developer tools such as Cursor. The address of Marvin's MCP server is the same for every compatible client; only the OAuth connection changes depending on the account used.

### Is MCP only useful for sourcing?

No. Sourcing is only one of the possible uses. Marvin's MCP server also covers CRM, ATS, outreach, assignments and timesheets, messaging and reporting: the actions a recruiter already performs in Marvin.
