Scrupp MCP server

Scrupp MCP server

An MCP server over the Scrupp Jobs API, so an agent can pull leads without anyone writing HTTP calls. Ships as scrupp-mcp-server on npm and as a listing in the MCP registry; that registry is what a growing number of agent runtimes read when they pick a tool, which is why this is the channel where the buyer is not a person.

In Claude (no install)

Add https://mcp.scrupp.com/sn as a custom connector in Claude and sign in to Scrupp. Setup steps, what each tool needs and troubleshooting: https://scrupp.com/docs/api/claude

In Claude Code (plugin)

/plugin install scrupp --marketplace Wism-Corp/scrupp-mcp-server

On Claude Code older than 2.1.275, in two steps:

/plugin marketplace add Wism-Corp/scrupp-mcp-server
/plugin install scrupp@scrupp

The plugin connects Claude Code to the hosted connector at https://mcp.scrupp.com/sn, the same full tool set as the private Claude connector, so nothing is installed locally. The first time a tool runs, open /mcp, pick scrupp and sign in to Scrupp.

The plugin carries two things:

first, pick the tool from what the user has, export five records as a sample before the whole search, collect long jobs with scruppgetjob, and what to say when a tool refuses.

• .mcp.json — the hosted connector above, signed in with OAuth.
• skills/scrupp-prospecting — how Claude should use the tools: check credits

Check it before a release with claude plugin validate ..

Using it

{
  "mcpServers": {
    "scrupp": {
      "command": "npx",
      "args": ["-y", "scrupp-mcp-server"],
      "env": { "SCRUPP_API_KEY": "sk_live_..." }
    }
  }
}

SCRUPPWAITSECONDS (default 120) caps how long a tool blocks. A job that outruns it comes back as { status: "running", jobid } with a note telling the agent to call scruppget_job — better than holding a conversation open for twenty minutes on a 2,500-record export.

Tools

Every tool carries a title and readOnlyHint / destructiveHint annotations — the Claude connectors directory rejects a server without them.

Hosted connector (Streamable HTTP)

node src/http.js (or scrupp-mcp-http) serves the same server over Streamable HTTP at POST /mcp, stateless, with GET /health for a load balancer. PORT defaults to

Authorization: Bearer sklive...; the key is scoped to that request, so one process serves many accounts without mixing them up.

1. Each request authenticates with the caller's own key as

The hosted connector exposes a narrower tool set than the local server: email finding, email verification, job status and credits. The tools that gather data from LinkedIn (Sales Navigator, LinkedIn search, profile enrichment) or answer from the shared lead base (company lookup, decision makers), and the Apollo export, stay local-only — they run through accounts that are not the caller's, which is not something to put in a public directory listing.

Try it in Claude Code against a deployed instance:

claude mcp add --transport http scrupp https://mcp.scrupp.com/mcp \
  --header "Authorization: Bearer sk_live_..."

Private connector with the full API (/sn)

Shortened. The full README is on GitHub.