MCP for recruiters: what the Model Context Protocol does and which recruiting tools support it

MCP, the Model Context Protocol, is an open standard that lets AI assistants such as Claude and ChatGPT connect to other software. For recruiters, it means asking an assistant questions about live data in your ATS, sourcing tool or interview platform, such as which finalists still lack a scorecard, and in some cases letting it act, such as adding a note, without exporting a spreadsheet. As of October 2026, Greenhouse, Ashby, Gem, BrightHire and Metaview publish official MCP servers. Access generally follows each user's existing permissions, some servers can write as well as read, and the data you pull into an assistant is then covered by the AI provider's terms.

What is MCP, in plain terms?

The protocol's own introduction calls MCP "an open-source standard for connecting AI applications to external systems" and compares it to "a USB-C port for AI applications": one standard plug instead of a custom integration for every pairing. Its architecture overview names three parts:

A server can offer three kinds of thing: tools, "executable functions that AI applications can invoke to perform actions"; resources, data that provides context; and prompts, reusable templates. For a recruiter, the tools are what matter: "search candidates," "get scorecards," "add note," "move stage." Servers run either locally on your computer or remotely on the vendor's platform; the recruiting servers below are remote, so you connect to a web address and sign in.

MCP itself makes nothing more accurate. It moves data to the assistant. The assistant can still misread it, summarize it wrongly or invent a detail, so the checking habits you use for any AI output still apply.

What can a recruiter actually do with MCP?

The value is in questions that cut across many records, which ATS reports often make slow. Examples of the kind vendors describe, phrased as a recruiter would ask:

• Pipeline status: "Which candidates in the final stage for the controller role have no scorecard from the hiring manager yet?"
• Interview evidence: "Across this week's screens for the data engineer role, what did candidates say about notice periods and salary expectations?"
• Interviewer patterns: "Which interviewers' scores most often disagree with the final decision?"
• Funnel analysis: "How has pass-through from phone screen to onsite changed for engineering roles over six months?"
• Admin: "Add a note to each of these five candidates saying the role is on hold," where the server allows writes.

The questions to avoid are the ones that turn the assistant into the decision-maker, such as "rank these candidates and tell me who to reject." That changes the legal picture, covered below.

Which recruiting tools offer an MCP server?

From each vendor's own documentation, as of October 2026:

Third-party and community MCP servers also exist for several ATSs, usually running on an API key. They are not the vendor's product, they often carry broader access than you need, and the vendor will not support them.

How do you connect an MCP server to Claude or ChatGPT?

Each vendor's setup guide has the exact server address and steps. The general shape in Claude, from Anthropic's custom connectors article:

1. Check your plan. Custom connectors work on Free, Pro, Max, Team and Enterprise plans, with Free limited to one custom connector. On Team plans, only owners and primary owners can add them for the organization.
2. Get the server address from your vendor's guide, and any admin set-up the vendor requires first, such as Greenhouse's scope selection.
3. In Claude's settings, open Connectors and add a custom connector with that address.
4. Sign in to the recruiting tool when prompted and approve the access level.
5. Enable the connector in a conversation and review each tool approval request before allowing it.

ChatGPT also supports MCP; OpenAI's MCP documentation and the vendor's own guide cover its steps. Anthropic's article adds a warning worth repeating to every recruiter: only connect to servers built and hosted by organizations you trust.

What are the risks of using MCP with candidate data?

Does using MCP change your legal obligations?

MCP is plumbing; the law looks at what the assistant does with candidate data. Summarizing scorecards for a recruiter who then decides is one thing. Asking an assistant to score, rank or filter candidates and acting on its output without real review can make it an automated employment decision tool under New York City's Local Law 144, or automated decisionmaking technology under California's rules from January 1, 2027; see CCPA ADMT rules for employers. Under UK and EU data protection, sending candidate data to a new AI provider is a new processing activity that your privacy notice and processor contracts should cover. An organization-wide view of approved tools belongs in an AI governance policy.

A checklist before you switch it on

• Use the vendor's official MCP server and its documented address.
• Have an admin set the narrowest scopes; begin read-only.
• Connect only from an AI plan approved for candidate data.
• Keep tool approval prompts on for anything that writes.
• Agree which questions are in bounds: job-related questions yes, anything about protected characteristics no.
• Tell recruiters to check figures and quotes against the source record.
• Record the connection in your AI register and review it when the vendor adds tools.

If you are not ready for live connections, the same questions can be asked of exported data with careful prompts; see Claude prompts for recruiters and the data rules in how to use AI in recruiting responsibly.