MCP server

ChurnLens runs a Model Context Protocol server, so an AI agent can compute retention, concentration and unit-economics figures directly. No key, no account, no rate limit, and nothing you send is stored.

Connect

Endpoint: https://churnlens.site/api/mcp (streamable HTTP)

Claude Desktop or Claude Code

npx mcp-remote https://churnlens.site/api/mcp

Any MCP client, by config

{
  "mcpServers": {
    "churnlens": {
      "command": "npx",
      "args": ["mcp-remote", "https://churnlens.site/api/mcp"]
    }
  }
}

A GET on the endpoint returns the manifest; the machine-readable descriptor lives at /.well-known/mcp.json.

Tools

ToolWhat it returns
calculate_churn_rateNRR, GRR, revenue churn, correctly compounded annualised churn, and the NRR−GRR spread that exposes churn masked by expansion
analyze_revenue_concentrationHerfindahl-Hirschman Index, top-N share, and which customers are large enough that losing one is a balance-sheet event
detect_zombie_mrrDormant-but-paying accounts and the ARR at risk behind them
score_saas_healthComposite 0–100 across five dimensions, plus the weakest one
calculate_ltvGross-margin-adjusted lifetime value, LTV:CAC and CAC payback
get_scoring_bandsEvery threshold the tools apply, with its provenance

Example

calculate_churn_rate({
  starting_mrr:    100000,
  expansion_mrr:    37000,
  contraction_mrr:   5000,
  churned_mrr:      17000
})

→ nrr_pct: 115, grr_pct: 78, expansion_masking_spread_pts: 37
  "A wide NRR-GRR spread: expansion revenue is masking substantial
   churn underneath. Diligence should look at the retained base
   separately from upsell."

Every tool returns structuredContent alongside the text block, so an agent gets typed numbers rather than prose it has to parse back out.

What it does not do

These tools compute from summary figures you supply. They cannot see the patterns that only emerge from customer-level data — cohort decay curves, renewal-cliff timing, concentration in specific logos. That needs the full monthly MRR-by-customer ledger, which is what our paid analysis ingests. Treat the MCP tools as a first-pass screen.

The same maths is open source under MIT, so an agent's output here can be reproduced and checked independently.

Related

Related