Korture Hiring MCP

Korture Hiring is a remote MCP server for hiring managers. Tell it what you want the person to get done, what the team is like, and how the place runs. It reads the role back to you and answers the question you actually have: who is likely to be wired for this, and are there enough of them.

It reads a role through four plain facets:

The work. What the job is at its core: hands-on, solving hard problems, or creating.
The team. How much of it runs through other people.
The drive. How much leading and persuading it asks for.
The culture. Whether it runs on structure and exactness, or loose and fast.

Each facet says where the role sits and where people suited to this kind of work usually sit, on one scale drawn from a real population. Where the role asks for more than those people normally bring, that is the rare part, and it is the reason a search stalls.

Then it counts. About 1 in how many people are wired for this role, measured against real people rather than against other job titles. If that number is small, it shows you what each make-or-break part is costing you: take this one off and about 1 in 7 becomes about 1 in 5. The preview is honest, so it will also tell you when a change makes the role narrower, or changes nothing.

When the role is already reachable, it moves you on: mint the scan link for candidates, and pull the structured interview for the role.

Free to use with a Korture account. Sign in at korture.com.

Endpoint: https://mcp.korture.com/api/mcp

Available tools

Tool	Purpose
`read_the_role`	Is what you have said enough to read the role from, what is missing, and what to go and ask. Call it before `create_brief`.
`create_brief`	Read the role into the nine behavioural demands: how critical each is, which are make or break, the wording each came from, and a "Will they ___?" question per hard part.
`get_brief`	The finished brief. Also the tool you poll while one is generating.
`read_the_people`	The four facets, about 1 in how many people are wired for the role, and what each make-or-break part would do to that if it came off.
`get_playbook`	The structured interview for the role: competencies from this job, questions, and what a strong, mixed or weak answer sounds like.
`start_scan`	The link a candidate opens to do their scan.
`get_scan_result`	The read on one candidate.

What it costs

A credit covers one of three things: a new role, a change to a role, or a candidate finishing a scan. Your first role is free. Reading anything, a brief, a role read, an interview or a candidate, is always free.

Running out of credits stops you adding a role. It never stops a candidate mid-scan and never puts a result behind a paywall.

What the numbers are not

The band on a candidate says how congruent that person is with this role. It is not a grade on the person and not a ranking against other candidates. Different never means worse. The tools return the read in words for that reason, and an assistant reading them out should not turn them into a score.

Connecting from Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

config.json

{
  "mcpServers": {
    "korture-hiring": {
      "url": "https://mcp.korture.com/api/mcp"
    }
  }
}

Quit Claude Desktop fully and reopen. The connector appears in your tray.

First time you connect

The connector requires a Korture account. After you add it, Claude Desktop shows a Connect button on its connectors page. Click it. You will be sent to hire.korture.com to authenticate via email OTP, no password. Authentication takes about 20 seconds.

Once you finish, return to Claude Desktop. All seven tools become available. Your account is remembered until you revoke it from the connectors page.

After you have a brief

Every brief-shaped response carries a links block:

links.overview: the brief on korture.com
links.role_shape: the four facets, and where a change to the role is actually applied
links.candidates: candidates and their results
links.playbook: the structured interview

The assistant will surface these at the end of a response. read_the_people tells you what a change would do; applying it is a decision, so it happens on the role shape page rather than in a chat.

If you previously connected to https://korture-mcp-server.vercel.app/api/mcp, update the URL above and re-authenticate. The old URL still serves traffic via a redirect, but the canonical address from now on is mcp.korture.com.

Connecting from Cursor (one click)

Click the button above in a browser that has Cursor installed. Cursor opens with the server name and URL pre-filled, you confirm and save.

Connecting from ChatGPT and other MCP clients

Any MCP client that takes a remote Streamable HTTP server URL works with the same address: https://mcp.korture.com/api/mcp.

ChatGPT. Add it as a custom connector. That needs developer mode on a paid plan.
Gemini Enterprise. Add an MCP server under connected apps, with the URL and your OAuth details.
Cursor. The one-click button above.

The server implements the 2026-07-28 protocol and serves 2025-era clients from the same endpoint, so it works whichever revision your client speaks.

Authentication

The server participates in OAuth 2.1 with PKCE (RFC 6749 + 7636, per the MCP spec). It advertises:

https://mcp.korture.com/.well-known/oauth-protected-resource: resource metadata (RFC 9728)
https://hire.korture.com/.well-known/oauth-authorization-server: authorization server metadata (RFC 8414)
Dynamic Client Registration at https://hire.korture.com/api/oauth/register (RFC 7591)

MCP clients that implement DCR register themselves automatically. For service-to-service callers, an internal-tier API key is available on request.

Rate limits are tier-based. Internal and admin tiers are unlimited; the free tier is 10 briefs per day. The MCP server does not accept unauthenticated calls.

Source and deployment

Code: api/mcp.ts, a thin stateless gateway to hire.korture.com/api/v1/*
SDK: @modelcontextprotocol/server v2 (the 2026-07-28 specification)
Runtime: Vercel Functions, Streamable HTTP, stateless
License: Apache-2.0

The MCP server holds no scoring logic, no prompts, and no demand definitions. Every tool forwards to the same REST API the product itself uses, so the answers here and the answers on screen cannot drift apart. That also makes this repo safe to fork or audit publicly.

Contributing

Issues and PRs welcome. For security disclosures, email team@korture.com rather than opening a public issue.