EasyHire Jobs — MCP server

Ask an AI assistant about jobs and get real EasyHire openings back, each with a link to its public page.

This is a Model Context Protocol server over the public EasyHire job board. It is read-only, holds no credentials, and can see nothing a candidate browsing the board could not.

Connect

The server is remote — there is nothing to install.

Claude Desktop / claude.ai — add a custom connector with the URL:

https://mcp.easyhire.me/mcp

Claude Code

claude mcp add --transport http easyhire-jobs https://mcp.easyhire.me/mcp

Any MCP client that takes a config file

{
  "mcpServers": {
    "easyhire-jobs": {
      "type": "http",
      "url": "https://mcp.easyhire.me/mcp"
    }
  }
}

No account, no API key, no sign-in. If a client asks for authentication, it is misconfigured — this server does not use any.

Related MCP server: trackly-cli

What you can ask

Capabilities

Each opening carries a url to its public page. Applying happens there.

What this deliberately does not do

Knowing the edges matters as much as knowing the features:

• It cannot apply for you. No applications, no messages, no uploads. It returns the link and stops.

• It shows only what is already public. Approved, open, public postings — the same set the board's own search returns. Nothing here can reach an unapproved, hidden, closed or sample posting, because visibility is decided entirely by the board and this server implements none of it.

• It returns no personal data. No hiring manager, recruiter or contact person ever appears in a response, even though the upstream data contains them. The shaping layer builds every field explicitly rather than passing the upstream object through.

• It does not inflate view counts. Retrieving a posting through an assistant does not count as a human viewing it — employers' view counts stay a measure of real interest.

• It is not employer-facing. No candidate pipelines, no interviews, no team data. Those are a different specification, not an extension of this one.

Two-step by design

search_jobs returns a short excerpt, because the board truncates the description before this server ever sees it. That excerpt is enough to decide which openings are worth reading and not enough to describe a role or judge a fit — get_job is for that. The capability descriptions tell assistants so, which is why a well-behaved assistant fetches details before making claims.

Development

Requires Node 22 (.nvmrc).

yarn install
yarn test          # unit tests for the shaping layer
yarn typecheck
yarn lint

cp .env.example .env   # point at non-production
yarn dev               # stdio, for a local client
yarn inspect           # stdio, through the MCP Inspector
yarn dev:http          # http://localhost:3000/mcp

yarn eval runs the evaluation suite against a deployed endpoint — the connector attaches a server by URL, so it cannot be pointed at a local process:

EASYHIRE_MCP_URL=https://mcp2.easyhire.me/mcp yarn eval

Configuration

Two values, both non-sensitive (see .env.example): the board API to read from and the public host used to build links. No secret belongs in this module — its runtime role grants nothing but logs, and that is a property worth keeping.

The one rule

mcp/ imports nothing from the monolith — not from core/, not from docker/, not from config/, and no types via @interfaces. Upstream shapes are re-declared locally in src/api/types.ts from the observed HTTP contract.

This is enforced by a dependency-cruiser rule (yarn depcruise:mcp from the repository root), not by intention. Breaking it would cost three things at once: the guarantee that this server sees only public data, its independent deploy, and the ability to extract it into its own repository.