Qampi MCP server

Make LinkedIn a step in your pipeline. This exposes Qampi's outreach engine as MCP tools, so an agent in Claude Code, Codex, Cursor — or any MCP client — can research a prospect, add them to Qampi, enroll them in a sequence, and read back what happened, without leaving the editor.

It is a thin client over Qampi's public REST API. It holds no LinkedIn credentials: your Qampi API key authenticates, and every safety rule (daily and weekly caps, the invite ramp, action pacing, the outstanding-invite ceiling) is enforced server-side. An agent cannot talk its way past a cap by phrasing a tool call differently.

Setup

1. Create an API key in Qampi → Settings → API keys (shown once).
2. Add the server to your client.

The quickest route in Claude Code or Codex is the plugin, which also teaches the agent the safe order of work. See qampi-plugins.

If your client doesn't pass environment variables through, save the key in ~/.config/qampi/api-key instead (or point QAMPI_API_KEY_FILE at a file).

Claude Code

claude mcp add qampi --env QAMPI_API_KEY=qampi_live_… -- npx -y @qampi/mcp-server

or commit .mcp.json to share it with your team:

{
  "mcpServers": {
    "qampi": {
      "command": "npx",
      "args": ["-y", "@qampi/mcp-server"],
      "env": { "QAMPI_API_KEY": "qampi_live_…" }
    }
  }
}

Cursor

.cursor/mcp.json (project) or ~/.cursor/mcp.json (global) — same shape as above.

Codex

~/.codex/config.toml:

[mcp_servers.qampi]
command = "npx"
args = ["-y", "@qampi/mcp-server"]
env = { QAMPI_API_KEY = "qampi_live_…" }

Modes — read this before enabling writes

| QAMPI_MODE | Behaviour | |---|---| | read-only (default) | Only read tools are registered. The agent cannot see, let alone call, anything that changes state. | | full | Adds leads, campaigns, enrollment, launch, email lookup, webhooks. |

Write tools are withheld rather than merely discouraged, because "the model was told not to" is not a safety property. Start in read-only; switch to full when you actually want the pipeline to act.

qampi_launch_campaign is the only tool that contacts real people. It is annotated destructive, and requires confirm: true — meant to represent a human's explicit go-ahead, not the model's own enthusiasm. A sent invitation cannot be unsent, and a withdrawn one cannot be resent for three weeks.

Tools

Read (always available)

| Tool | What it does | |---|---| | qampi_status | LinkedIn connection + health, plan, and today's remaining budget. Call this first — the invite allowance ramps with account age and acceptance rate, so it is not a fixed number. | | qampi_list_templates | Prebuilt sequences. | | qampi_list_leads / qampi_get_lead | Leads and their enrichment. | | qampi_list_campaigns / qampi_get_campaign / qampi_campaign_leads | Campaigns and per-lead progress. | | qampi_search_people | LinkedIn people search. Costs one search from the monthly commercial-use budget. |

Write (QAMPI_MODE=full)

| Tool | What it does | |---|---| | qampi_add_leads | Adds leads. Contacts nobody. | | qampi_annotate_lead | Tags and notes only — enrichment is engine-written, status is engine-derived. | | qampi_create_campaign | Creates a campaign in DRAFT. | | qampi_enroll_leads | Adds leads to a campaign. Still sends nothing. | | qampi_launch_campaign | Starts real outreach. Requires confirm: true. | | qampi_find_email | Finds a work email. Consumes one credit. | | qampi_register_webhook | Subscribe a URL to lifecycle events so a pipeline reacts instead of polling. |

Environment

| Variable | Default | | |---|---|---| | QAMPI_API_KEY | — | Required, unless the key file below exists. | | QAMPI_API_KEY_FILE | ~/.config/qampi/api-key | Read when QAMPI_API_KEY is unset. | | QAMPI_MODE | read-only | full enables write tools. | | QAMPI_BASE_URL | https://api.qampi.com/api/public/v1 | Point at a staging API. | | QAMPI_TIMEOUT_MS | 60000 | Searches can take a while. |

Example

"Find the three companies we mentioned in notes.md, look up their heads of growth on LinkedIn, and add the 2nd-degree ones to Qampi tagged q4-pilot."

The agent reads your file, calls qampi_search_people, filters by degree, then qampi_add_leads + qampi_annotate_lead. Nothing is sent to anyone — launching stays a separate, deliberate step.

Development

npm run build      # esbuild → dist/index.js
npm run smoke      # drives the server over stdio against a live account

npm run smoke needs QAMPI_API_KEY and makes real (read-only) API calls.