paper · Growth Funnel

How to read this page

The premise

Seats are distribution. Agent usage is the paywall. Free gives every team unlimited editors and 100 MCP calls a week. Pro charges $20 per editor and opens the cap to 1M calls. So the business does not convert on seats; it converts on agent behavior. This page is instrumented accordingly: it counts agent connections, agent edits, tool calls, and the moment a team touches the cap. Almost nothing here counts logins or canvas opens.

Reading order

• Summary, WatchThe week in five points and the three things that need an owner. Every number in them is a link to its source.
• 00 HeadlineThe section title is the year-end pace, computed live: conservative (flat net-new, no compounding) against the $3M target, with the compounding case beside it. Two minutes, in silence, before anyone speaks.
• 01 FunnelWhere teams fall out between signup and paid, on a cohort old enough for every stage to have happened. The stage map underneath shows which raw event means which stage.
• 02 The paid wallIs the 100-call cap doing its job? Cap-hit conversion is the load-bearing number of the pricing model.
• 03 EngagementIs the agent habit deepening, and which paying teams have gone quiet? Leading indicator for both conversion and churn.
• 04 Revenue qualityDoes the paying base grow on its own, and is the launch wave durable? Momentum and durability are never blended.
• 05 LoopsWhich acquisition loop gets funded, and what agents fail at, which is what we should document next.
• 06 OrganizationsWhich companies are already an Organizations deal and don't know it. There is no sales team; this view is the sales team.
• 07 Unit economicsHow much room the $20 seat has, and what free costs. The customer brings their own agent tokens; image generation is our variable cost.

Interactions

• ⓘOpens the card's governed definition: what it is, the formula, the source tables, the SQL it resolves to, the decision it informs, and its caveats. Definitions live in metrics.yml; the file itself opens from the header or the side nav.
• Click a cardTraces the number back to its source: one real team followed from the raw event stream through the model tables to the exact cell on the card. Highlighted column is the field the metric reads; highlighted row is the one that contributed.
• Click a numberAny bold number in the Executive Summary or Watch Areas traces the same way.
• [−] / [+]Collapses the Summary and Watch cards for the meeting.
• Hover a chartShows the exact value at that week, every series on the chart, and the total where there is one. Projected points are labelled as forecast.
• ProjectionExtends every line chart to Dec 31 as a dotted line past a vertical marker, and adds a chip to the card with the projected value. Level metrics (ARR, paying teams, active teams) compound at the chosen basis: the trailing 8-week rate, or the rate since the post-launch baseline began on Apr 13 (the launch spike itself is excluded, because compounding from it produces nonsense); rates are held flat at their trailing 4-week average, because a rate does not compound. The actual numbers on the cards never change.
• escCloses any panel.

Conventions

• Mature cohortsRates that need time to happen (activation in 7 days, cap-hit conversion in 30) are shown only for cohorts old enough to measure. Immature cohorts are dotted and must not be read as a trend.
• Week statusThe badge next to the week says whether the week is complete or partial. Weekly deltas compare complete weeks only.
• WavesTeams are tagged by the launch wave they arrived in (desktop launch, Config, YC review). Retention is reported by wave and never as one blended curve.
• BluePaper's accent. New MRR, the cap, and anything the page is pointing at.
• EmeraldExpansion and positive deltas.
• YellowContraction.
• Light redChurn.

The data

Everything here is synthetic, generated to match public figures and the numbers Stephen shared: roughly $46K ARR at the March desktop launch, $1.8M in mid-August, 25% growth in August. Customer logos from paper.design appear as domains; every number attached to them is invented. The pipeline, the semantic layer, and every definition are real and would run unchanged on live Stripe, PostHog, and MCP server data.

Sources