Prompt your pipeline

Connect your own AI assistant to HookGet over MCP and run the whole product by conversation: wire sources, build and edit dashboards, read the numbers, diagnose failures. Thirty-five tools, five expert playbooks, one scoped key — and every prompt runs on your LLM, not ours.

In short

npx @hookget/mcp with a scoped API key — the dashboard mints an "AI agent" preset key in one click.
35 tools cover the whole surface: sources, dashboards, the data advisor, events, endpoints, diagnostics.
Five playbooks ship as MCP prompts: set-up-my-funnel, build-my-dashboards, weekly-report, diagnose-my-pipeline, optimize-my-dashboards.
The deterministic layer does the heavy thinking server-side (advisor, templates, validation) — your model narrates and confirms, which is what keeps token cost small.
Safety is structural: scoped key, whole-document validation, no tenant vocabulary, confirm-gates on anything destructive.

The setup, complete

# 1. Dashboard → Keys → Create key → preset "AI agent (MCP)"
# 2. In your MCP client config:
{
  "mcpServers": {
    "hookget": {
      "command": "npx",
      "args": ["-y", "@hookget/mcp"],
      "env": { "HOOKGET_API_KEY": "hg_live_…" }
    }
  }
}
# 3. Talk: "set up my marketing funnel" — the playbook takes it from there

What each side does

Your LLM (your account)	HookGet (deterministic, no LLM)
reads the playbook, narrates, asks you	serves 35 validated, tenant-scoped tools
chooses what to propose	advisor precomputes proposals with reasoning and evidence
sends a dashboard document	validates it whole; a document cannot name a tenant
asks for numbers	compiles and runs the queries, returns JSON
pays for its own tokens	adds zero AI cost to your bill

The playbooks

Playbook	What your assistant does
`setup-my-funnel`	inspects what flows, names the missing funnel legs, connects sources with your credentials, finishes with the command center
`build-my-dashboards`	one advisor call → proposals with reasoning → creates what you approve
`weekly-report`	reads every dashboard and writes the 300-word report a marketing lead would send — strictly read-only
`diagnose-my-pipeline`	sweeps sources, stats and dead letters; one likely cause, its evidence, one proposed action
`optimize-my-dashboards`	finds dead widgets, unshown data families and missing formulas; edits only on approval

Questions

Whose LLM does the thinking, and who pays for it?

Yours, and you — which is the point. HookGet ships no model and never sees your prompts: your assistant (Claude, or anything MCP-capable) connects with a scoped HookGet key and does the reasoning on the account you already have. Our side is deterministic — the advisor, the validator, the compiler — so the expensive thinking is precomputed server-side for free, and the playbooks keep your model's share short and directed. Token cost stays where it belongs, and stays small by design.

How is an AI agent with write access safe?

The agent holds a key with exactly the scopes you minted (the dashboard has a one-click "AI agent" preset: data reads, sources, dashboards — never key minting, members, or publishing). Every write passes the same whole-document validation as a human click; a dashboard cannot name a tenant; destructive tools demand an explicit confirm. The model proposes — the validated layer decides. A confused or hostile prompt produces, at worst, a refused document.

Can it work alongside my other MCP servers?

That is the intended shape. Your assistant can hold HookGet's MCP next to your CRM's, your repo's, your automation platform's — and orchestrate across them: read a number from a HookGet dashboard, open a ticket elsewhere, trigger an n8n flow that reports back into HookGet as events. Each server enforces its own permissions; HookGet's key only ever opens HookGet.

What are the playbooks?

Five expert procedures shipped inside the server as MCP prompts — set up the funnel, build dashboards, weekly report, diagnose the pipeline, optimise dashboards. Your model follows a tested recipe instead of improvising: fewer tokens, repeatable results, and the safety rules re-read on every run.