In the last 60 days, five of the biggest names in recruiting software shipped or announced an MCP server. Workable went live on May 13 with 38 tools, free on every plan. Greenhouse announced its own on May 7 and started rolling it out in June. Gem previewed GeMCP on June 23, HeyMilo shipped theirs a day later, and Loxo's CEO announced Loxo MCP in early July.
If those names cover any part of your stack, something changed for you this spring: the AI assistant you already pay for can now read your pipeline, search your candidate database, and draft your reports. Live, from the source, without a single CSV export.
Most recruiters haven't connected any of it. This guide covers what MCP is in plain terms, which recruiting tools have a server today, five workflows worth setting up first, and what to check before you give an AI assistant write access to your ATS.
What MCP is, in recruiter terms
MCP stands for Model Context Protocol, an open standard Anthropic released in November 2024 and later handed to the Linux Foundation's Agentic AI Foundation. Think of it as a USB port for AI: one plug shape that lets any assistant connect to any tool that supports it. Claude, ChatGPT, and Microsoft Copilot all speak it.
Before MCP, getting your ATS data in front of an AI assistant meant exporting a CSV, pasting it into a chat window, and repeating that ritual every time the data moved. An MCP server removes the middle step. The vendor exposes a set of named tools ("search candidates," "list open jobs," "update stage") and your assistant calls them directly, with your permissions, in the middle of a conversation.
The practical difference is one sentence long. You type "show me every backend candidate who has been sitting in stage two for more than 14 days" and get an answer from your live ATS, not from a stale export.
Who has an MCP server, as of July 2026
This table covers official, vendor-supported servers. The ecosystem moves monthly (three of the entries below are less than six weeks old), so treat the date in this heading as part of the data.
| Tool | Status | What it gives your assistant | Cost |
|---|---|---|---|
| Workable | Live (May 2026) | 38 tools: jobs, candidates, stages, offers, plus HR data | Free on all plans |
| Greenhouse | Rolling out (June 2026) | 36 tools: pipeline analysis, summaries, governed write access | Included |
| SeekOut | Live | Search 1B+ profiles from Claude, ChatGPT, or Copilot | Included |
| Manatal | Live (late 2025, first ATS to ship) | Candidate search, notes, outreach drafts, summaries | Enterprise Plus plan |
| Zoho Recruit | Live | Candidates, pipelines, hiring workflows | Included |
| Pinpoint | Live | Query hiring data, move candidates, draft outreach | Included |
| Recruitly | Live | Candidates, pipelines, revenue reports | Included |
| Leonar | Live | Sourcing across 870M+ profiles, outreach | Included |
| Pin | Live | Ten tools: search, review, outreach kickoff | Free tier |
| Checkr | Live | Background check status, with PII stripped before the AI sees it | Included |
| Indeed | Beta | Job search and listings data | Not disclosed |
| HeyMilo | Live (June 2026) | AI interview data, scheduling | Included |
| Loxo | Announced (July 2026) | Claude or ChatGPT straight into your private Loxo database | TBA |
| Gem | Announced (June 2026) | Full-funnel recruiting data, first look published | TBA |
Just as telling is who's absent. LinkedIn has no official MCP server. Neither do Juicebox, hireEZ, iCIMS, Phenom, or Paradox. For the biggest database in recruiting, the only MCP access is community-built scrapers that violate LinkedIn's terms of service and put your account at risk. More on that below.
A note on community servers. Developers have published unofficial MCP servers for Ashby (67 tools), Lever (59 tools), Bullhorn, and dozens more, usually as open source on GitHub. Some are excellent and expose more of the platform's API than official servers do. They are also one developer's side project holding your candidate data, with no vendor support and no guarantee of updates. Fine for experimenting on a test account. For production pipeline data, wait for the official server or ask your vendor when one ships.
Five workflows that work today
Everything below runs in Claude or ChatGPT once the relevant server is connected. No code, no Zapier chains. Each workflow names the prompt pattern and the thing to double-check.
1. Ask your ATS questions in plain English
The most immediate payoff. Once your ATS server is connected, questions you used to answer with filters, exports, and a pivot table become a single message: "Which of my open reqs had no candidate activity this week?" or "Summarize where every finalist for the platform engineer role stands."
Greenhouse built its early rollout around exactly this. Its design partners used the MCP to build Slack copilots that answer hiring managers' status questions directly, so the recruiter stops being a human lookup service. Ask for the summary before your Monday sync instead of building it the night before.
What to check: the first few times, verify the numbers against the ATS interface. Assistants occasionally miscount when a query spans stages or date ranges, and you want to know your server's weak spots before you quote its output in a leadership meeting.
2. Turn an intake call into a search brief
This workflow chains the assistant's native strengths (reading a transcript, structuring a brief) with MCP access to your tools. Drop the intake call transcript into the chat, have the assistant draft the JD and a sourcing brief with must-haves and deal-breakers, then let it check the brief against reality: open roles already in your ATS, similar past searches, candidates already in your database who match.
We've covered the transcript-to-brief pattern in detail in our Claude for recruiters and ChatGPT for recruiters guides. MCP upgrades it from "draft me a brief" to "draft me a brief and tell me what we already have."
3. Source candidates without leaving the chat
With a sourcing server connected (SeekOut, Leonar, and Pin all offer one), "find senior Go engineers in Berlin with AWS experience" returns candidates inside the conversation. The assistant handles the phrasing; the server runs the search against its own index.
An honest caveat: the assistant contributes none of the search quality here. Results are only as good as the data layer underneath, its freshness, its coverage, and the signals it carries. A generic LLM plugged into a thin database gives you fluent access to weak results. That is the whole argument for purpose-built sourcing engines, and it's why the interesting question about any sourcing MCP is not "does it talk to Claude" but "what does its index know that others don't."
4. Wake up your own candidate database
The candidates you already interviewed are the cheapest sourcing channel you have, and the least used. With an ATS or CRM server connected: "Find every fintech product manager we spoke with in the last 12 months who reached at least stage two, and draft a re-engagement email for each that references the role they interviewed for."
That prompt used to be an afternoon of export, filter, and mail merge. It's now one message, and the draft emails cite real history instead of "I came across your profile." We wrote about why rediscovery beats fresh sourcing in candidate rediscovery; MCP removes its last practical barrier.
What to check: keep a human review step before anything sends. Drafting at scale is safe. Sending at scale without eyes on it is how you email a candidate you hired last quarter.
5. Pull reports without the Friday export ritual
Weekly pipeline digest, offer-stage summary, QBR prep: recurring reports are the workflow where MCP pays for its setup time fastest, because the cost repeats every week. Ask for the same report with the same structure each Friday, then paste it into your deck or your update email. Teams on Greenhouse's early rollout use exactly this pattern for board-ready hiring summaries.
If your reporting pain is bigger than one weekly digest, our guide to recruiting workflow automation covers where automation helps and where it quietly breaks things.
How to connect a server, in about 10 minutes
The setup is deliberately boring. In Claude, open Settings, then Connectors, then add the server. Official servers from the table above are either listed in the connectors directory or added by pasting the URL from the vendor's docs. You sign in to the vendor with your normal account, approve the access scopes, and the tools appear in your chat. Workable's version of this is a three-step help center article, and it's free on every plan, which makes it the easiest first test if you're on Workable. ChatGPT follows the same pattern through its connectors settings on paid plans.
One question worth asking any vendor before you connect: does the MCP server respect my user permissions? The good implementations do. Workable authenticates through OAuth2, so an assistant operating under your login sees only what your role sees. Greenhouse adds org-level controls, rate limits, and blocks on actions like extending an offer without human approval. If a vendor's answer is "the server uses one admin API key for everyone," that's a no.
What to check before you plug in
Four things separate a useful setup from a mess.
Official beats community, every time it exists. A community server with more tools is still an unmaintained dependency with access to candidate PII. The gap closes as vendors ship official servers; until yours does, experiment on test data only.
Leave LinkedIn alone. The community LinkedIn MCP servers work through browser automation, which LinkedIn's terms of service prohibit. Low-volume use may fly under the radar, but the account at risk is the one your business runs on. No workflow in this article is worth it.
The agent's actions are your actions. GDPR, CCPA, and EEOC exposure doesn't change because an assistant clicked the button. Give write access only where a human reviews the output before it lands (outreach drafts yes, auto-sent sequences no), and make sure your candidate privacy notice covers AI-assisted processing.
Date everything. The table above will be outdated within a quarter, and so will any vendor comparison you make internally. Re-check before you commit, and ask vendors without a server for their timeline; Forrester expects 30% of enterprise software vendors to ship MCP servers during 2026, so "not yet" often means "next quarter."
Where this goes next
The current wave is "chat with your ATS." The next one is agents that run whole workflows across several tools at once: read the intake transcript, search three sources, shortlist against your rubric, draft the outreach, log everything back to the ATS, and hand you the exceptions. The protocol plumbing for that future is what's being installed right now, tool by tool.
Glozo is building for that stack directly. Our Sourcing Agent already runs searches in the background on three signals a generic assistant can't reach: match rationale from the Skill Graph, Market Value, and the Open to Offers signal that surfaces passive candidates who are likely receptive. And in summer 2026 Glozo is launching its own MCP server, connecting Claude, ChatGPT, or whichever assistant you use straight to Glozo's search and market data. Your LLM runs the conversation; Glozo is the engine underneath it.
If you're deciding how agents fit your desk, start with our comparison of a purpose-built sourcing agent vs. a custom GPT and the checklist in how to choose an AI sourcing agent.
The recruiters who win the next two years won't be the ones with the most tools. They'll be the ones whose tools talk to each other, and MCP is how that conversation happens.