@growth-loop/mcp-server

The AI growth engineer for vibe-coded SaaS — talks to your Claude.

Model Context Protocol server that connects growth-loop.dev — funnels, retention, peer benchmarks, and the AI growth engineer — to Claude Desktop, Claude Code, and any other MCP-compatible client.

Once installed, Claude can answer questions like:

"What changed in my MRR last week?" "Why is my signup→paid funnel down?" "How does my D7 retention compare to other B2B SaaS at $1–10k MRR?"

…with answers cited to actual rows in your funnel — not a chat full of guesses.

Install

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent on your platform:

{
  "mcpServers": {
    "growth-loop": {
      "command": "npx",
      "args": ["-y", "@growth-loop/mcp-server"],
      "env": {
        "GROWTH_API_KEY": "pk_live_…",
        "GROWTH_HOST": "https://api.growth-loop.dev"
      }
    }
  }
}

Restart Claude Desktop. Open a chat, click the slider icon — growth-loop tools will be listed.

Claude Code

From inside your project:

claude mcp add growth-loop \
  -e GROWTH_API_KEY=pk_live_… \
  -e GROWTH_HOST=https://api.growth-loop.dev \
  -- npx -y @growth-loop/mcp-server

Auth

Two modes — pick one (or both):

| Env var | What it unlocks | |--------------------------|-------------------------------------------------| | GROWTH_API_KEY | Read endpoints (overview, funnels, retention, anomalies, commits, benchmarks). Get one at growth-loop.dev/dashboard/setup. | | GROWTH_SESSION_TOKEN | All of the above + growth_engineer_ask (the headline tool). Copy the value of the growth_session cookie from your browser devtools while logged in. |

GROWTH_HOST defaults to https://api.growth-loop.dev. Override for self-host. GROWTH_PROJECT_ID is optional — set if you want to pin a specific project.

What it exposes

Tools (12)

| Tool | Purpose | |-------------------------|--------------------------------------------------------------------------| | overview | Top-line snapshot: MRR, active users, signup→paid, deltas. | | metric_series | Time series for any metric. | | funnels_list | List configured funnels. | | funnel_get | Conversion + drop-off for a named funnel. | | retention | D1 / D7 / D30 cohort retention. | | diagnose | Closed-loop investigation on a metric — ranked hypotheses with citations.| | why_event | Explain why users drop on / interact with a specific event. | | anomalies_list | Recent statistical anomalies with candidate hypotheses. | | commits_list | Git-aware analytics — commits annotated with measured metric impact. | | peer_benchmark | Compare to anonymized peer SaaS at the same category + stage. | | hypotheses_list | RICE-ranked hypothesis backlog. | | growth_engineer_ask | The headline tool. Ask the AI growth engineer anything — grounded in your funnel, retention, peer benchmarks, hypotheses, deploys, and the last 30 raw events. |

Resources

• growth://overview
• growth://anomalies/recent
• growth://commits/recent
• growth://benchmarks/peers
• growth://hypotheses

Prompts (slash commands)

• /growth-init — auto-instrument this repo with the SDK and open a PR.
• /growth-weekly — what changed, biggest mover, 3 actions for the week.
• /growth-pmf — score 0–10 + what blocks the next click up.
• /growth-pivot — cohorts outperforming by ≥2× on retention or LTV.
• /growth-icp — who pays the most, retains the longest.
• /growth-strategy — 90-day plan, one focus area.
• /growth-diagnose <metric> — root-cause analysis for a dropped metric.

(Prompts are prefixed with growth- so they don't collide with Claude Code's built-in /init or other editors' built-ins.)

• /diagnose <metric> — root-cause cross-referencing events + commits.

Example session (Claude Code)

> /weekly

Last week (Apr 21–27 vs Apr 14–20):
  • MRR: $4,210 → $4,820 (+14.5%) [data_citations: subscription_created × 12]
  • signup→paid: 4.1% → 3.6% (−0.5pp)  ← biggest mover
  • D7 retention: 31% (cohort_size=84)

Top three actions:
  1. Diagnose the signup→paid drop. Two new commits to /onboarding/step-2
     landed Apr 23 — possible regression. (commits: 8a3f2c1, 9b1d4e2)
  2. Run a peer benchmark — D7=31% is below B2B-SaaS-0_1k_mrr p25=38%.
  3. Write up the +14.5% MRR week as your weekly digest's win.

Want me to open a hypothesis for any of these?

Self-host

Source: https://github.com/growth-loop/growth-loop.dev/tree/main/packages/mcp-server

The MCP server is MIT-licensed and ships as a single Node binary. If you self-host growth-loop.dev, point GROWTH_HOST at your own deployment.

Why this exists

We're building growth-loop.dev: an AI growth engineer for vibe-coded SaaS. Most product analytics tools assume you'll spend a quarter wiring taxonomies and weeks pasting CSVs into ChatGPT. We don't.

• Claude Code instruments the code. One claude /growth init and your funnel events ship with your next deploy.
• The growth engineer reasons about the data. Persistent threads, citations to real rows, weekly digest, hypothesis state machine.
• MCP makes Claude itself the interface. This package is how that works.

If you want the platform: https://growth-loop.dev. If you want to follow along: https://twitter.com/growthloop_dev.

License

MIT