This reads well and holds together. The piece is grounded in the two required citations (exact quotes preserved verbatim), stays honest about Lycato's current alpha/self-hosted-only status (verified against README.md and docs/hosted/README.md rather than asserting the intended £39/month hosted price as live), and matches the brand voice in BRAND.md (calm, declarative, diff/git mental model, no hype words). Final word count: 872 words in the body (942 total with frontmatter), within the 800-1100 target.

Here is the final post:

title: "BYO keys: per-seat AI pricing is a tax on curiosity" slug: "byo-keys-not-per-seat" dek: "Notion, Coda, and credit-metered agent tools all charge extra the moment you touch AI — Lycato doesn't, because there's no AI tier to buy into in the first place." audience: "Solo and small-team founders deciding whether their next tool's AI bill scales with headcount or with actual usage." tags: [pricing, byo-keys, open-source, self-hosted, agents]

Annika Helendi's review of Lindy names a side effect that rarely shows up on a pricing page: "I found myself avoiding experimenting because each interaction costs credits" (source). The product wasn't broken. The meter was working exactly as designed — it just made experimenting feel expensive enough to skip.

That's the real cost of most AI pricing today. Not the number on the plan page. The tax on curiosity: every test run, every "let's see what this does," is a debit against a balance you can watch fall. People adapt to that the same way they adapt to a parking meter — they leave early. It's a rational response to the incentive, not a character flaw, which is exactly why it's worth looking at the incentive instead of the person.

The bundled version of the same tax

Notion arrived at the same outcome from a different direction. AI used to be an optional $10-a-month add-on. It isn't optional anymore — it's folded into the base per-seat price, now $20 or more a seat. As eesel's breakdown of the change put it: "Notion killed the $10/month AI add-on, for a team of 10 that is at least $200/month." Nobody on that team asked for AI to get more expensive. The cost just moved from a line item someone could decline to a multiplier on headcount nobody controls.

Coda runs a version of the same idea through a different door. What's gated there isn't AI specifically — it's editing. Add a teammate who needs to actually touch a doc, not just read it, and you've added a paid seat. Different lever, same shape of bill: cost tracks who's on the roster, not what got used.

Credits or seats, the mechanism underneath is identical. Something is watching, and it prices your next click before you make it.

What changes when the key is yours

Lycato doesn't have a credit balance, and it doesn't have an AI seat, because Lycato never sits between you and the model provider. You bring your own key — Anthropic, OpenAI, Gemini, whichever you already have — and give it to your own machine:

lycato keys set ANTHROPIC_API_KEY sk-ant-...

From there, every agent call in your Lycato repo goes straight from your machine to the provider you chose, billed at that provider's own per-token rate. Nothing gets repackaged into a credit. Nothing gets marked up in transit — Lycato is AGPL-3.0 and self-hosted, so there's no billing layer of ours to insert even if we wanted one. If a run costs your provider nine cents, it costs you nine cents.

That changes what you're willing to try. Asking an agent to draft three versions of an investor update, one of which you'll actually use, costs what the tokens cost — not what a credit costs. You read the result as a git diff and merge the one you want. Throwing away the other two isn't waste against a shrinking allowance. It's just how the substrate is supposed to work: drafts are cheap, review is the only gate that matters, and nothing becomes real until you approve the diff.

What this doesn't solve

BYO keys removes the markup. It doesn't remove the bill. You are still paying a model provider real money for every token an agent uses, and Lycato doesn't cap it, forecast it, or interrupt a run to warn you — that ledger lives on your provider's dashboard, not ours. Leave an agent looping overnight on a bad prompt and it is still your card that clears.

It also means one more signup before anything works. You need an Anthropic, OpenAI, or Gemini account with billing attached, which is a real step — a tool that bundles AI in gives you a single invoice and nothing else to configure. BYO keys trades that convenience for a direct relationship with the provider, on purpose, but it is a trade, not a free upgrade.

Lycato also isn't zero-effort. Today it is a CLI and a local cockpit you run yourself, against a git repo you host — there is no managed version yet, which means secret storage, backups, and scheduling are your problem, not a support team's. That is a real cost, just a different kind: time and ops instead of a subscription line. A hosted convenience layer is on the roadmap, and the design intent for it is a flat infrastructure fee, not a per-token one — but that is a plan, not a shipped feature, and it is worth saying plainly rather than implying it already exists.

The actual trade-off

Per-seat and per-credit pricing are predictable for the vendor, not honest about usage. A raw token bill is the opposite: less predictable up front, because it moves with what you actually run instead of a flat figure a pricing page can print in bold. What you get in exchange is a cost model with exactly one input — tokens, at the provider's own rate — and no second meter charging you for finding out what the tool can do. Add a cofounder and the price doesn't change. Try something that doesn't work and the price doesn't change. Read the diff. Keep it or don't.