PumpGTM MCP server

Run LinkedIn, email and X outreach from any AI agent.
PumpGTM finds the buyers showing intent this week, drafts and sends the outreach from your own accounts inside each platform's limits, and hands every reply back to a human. This is the public reference for the hosted MCP server at https://mcp.pumpgtm.com/mcp.

Quickstart · Docs · tools.json · REST API · Platforms and agencies · YouTube

What your agent is driving

The same workspace the dashboard shows. Plays find people, sequences reach them across LinkedIn, email and X, and every reply waits for a human.

How it fits together

flowchart LR
  A[Your AI agent<br/>Claude, Cursor, Codex, your own] -- MCP over Streamable HTTP --> M[mcp.pumpgtm.com]
  M --> W[Your PumpGTM workspace]
  W -- find_people --> P[Buyers showing intent]
  W -- sequences --> L[LinkedIn]
  W -- sequences --> E[Email]
  W -- sequences --> X[X]
  L & E & X -- replies --> R[Human decides]
  R -- webhook reply.received --> A

Watch and read

What is PumpGTM, and how does it actually work? 4 minute walkthrough

Using AI to run founder-led outbound for 10 YC companies Namanyay's article on X: one lead universe, shared prospects, copy measured by meetings. Read on pumpgtm.com

More on the PumpGTM YouTube channel and in PumpGTM in the wild, videos founders and creators made on their own channels.

Connect

Endpoint (Streamable HTTP): https://mcp.pumpgtm.com/mcp

Two ways to authenticate, both scoped to one workspace:

• OAuth 2.1 for Claude (web and desktop), Claude Code, Codex, Cursor and ChatGPT. Add the endpoint and sign in; no key is needed. Dynamic client registration and PKCE are supported; metadata is at https://pumpgtm.com/.well-known/oauth-authorization-server. Access tokens last 30 days, refresh tokens 90.
• Workspace key for your own agent, the MCP SDKs, and the REST API. Sign in at app.pumpgtm.com, open MCP in the left nav, and copy the key. Send it as Authorization: Bearer eve_mcp_.. on every request. The key does not expire. Rotation is not self-serve yet: write to hello@pumpgtm.com and we rotate it for you.

The key's database row fixes the workspace. It cannot read or act on any other workspace.

import { Client } from "@modelcontextprotocol/sdk/client/index.js";
import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";

const client = new Client({ name: "your-agent", version: "1.0.0" });
await client.connect(
  new StreamableHTTPClientTransport(new URL("https://mcp.pumpgtm.com/mcp"), {
    requestInit: { headers: { Authorization: "Bearer YOUR_WORKSPACE_KEY" } },
  }),
);
const { tools } = await client.listTools(); // up to 26 tools, listed below

ChatGPT (until the PumpGTM listing is approved in the ChatGPT directory): in Settings, open Plugins and turn on Developer mode (Plus, Pro, Business or Enterprise). Click Plugins in the sidebar, then +: choose Add MCP Server in the desktop app, or Create app then Create MCP App in the browser. Name it PumpGTM, paste https://mcp.pumpgtm.com/mcp, keep OAuth, and click Create. Sign in to PumpGTM, allow the connection, then ask ChatGPT to find buyers or type @PumpGTM.

Claude Code: claude mcp add --transport http pumpgtm https://mcp.pumpgtm.com/mcp, then run /mcp (or claude mcp login pumpgtm) and sign in. Codex: codex mcp add pumpgtm --url https://mcp.pumpgtm.com/mcp then codex mcp login pumpgtm. Cursor: add a server with only the URL and sign in when it asks.

Install as a plugin

The PumpGTM plugin bundles this server with three skills: write a cold message, find buyers, and handle replies. In Claude Code:

/plugin marketplace add pumpgtm/pumpgtm-mcp
/plugin install pumpgtm@pumpgtm

Then run /mcp and sign in to PumpGTM. The plugin's manifest, skills and server config live in github.com/pumpgtm/pumpgtm-mcp.

In Codex:

codex plugin marketplace add pumpgtm/pumpgtm-mcp
codex plugin add pumpgtm@pumpgtm
codex mcp login pumpgtm

How results come back

• Every tool result is one MCP content block of type text whose text is a JSON string, and the same object in structuredContent. Parse the JSON.
• Most results include next: { tool, arguments, reason, requiresConfirmation, requiresNewRequestId? }. It is the suggested following call, not an instruction to skip the customer.
• Errors return isError: true and { status, retryable, changed, message, details }. changed: false means nothing was written. A missing object returns code resource_not_found.
• Tools that write take a requestId: 8 to 120 characters of letters, digits, . _ : -, starting with a letter or digit. It is an idempotency key. Use a fresh one per call; reuse one only to retry the same call.
• Metered workspaces carry Energy: one Energy is one person found. get_workspace returns energy with the monthly allowance, used and remaining, and find_people refuses with energy_exhausted when a run would exceed it. Workspaces without Energy see no such field. See Energy.
• Nothing reaches LinkedIn, email or X until a person approves it. find_people, review_people, save_sequence_draft and add_leads never send. Sending starts with set_sequence_status active on an approved sequence, and each reply is answered only through decide_reply.

A complete flow

1. get_workspace to read setup state, sequences, unfinished reviews and pending replies.
2. find_people with targeting and confirmed: false to save a Play and show the targeting summary. Then find_people with playId and confirmed: true to run it.
3. review_people with decisions and finish: true to queue the good fits into a sequence.
4. save_sequence_draft with steps, then set_sequence_status approved with expectedRevision, then active.
5. list_pending_replies, show the draft to the customer, decide_reply.

Reporting: get_workspace (view: "summary" for the funnel and account pacing, view: "activity" for the ledger) and list_leads with stage or sinceDays. The same numbers are available over REST at /docs/api.

Resources and prompts

resources/list returns six lead-search guides: gigacatalyst://lead-search/catalog, quickstart, fields, query-builder, patterns and results. Read fields before writing targeting by hand. One prompt, gigacatalyst_find_people, takes requestedCount ("10" to "100").

Versioning

Tools scheduled for removal say so in their description for at least 90 days before leaving tools/list. There is no fixed rate limit on the MCP server today; discovery is paced by each workspace's daily provider budget.