What it costs

What it costs to run

On your laptop it is free and there is nothing to sign up for. Deployed, it starts at about five dollars a month. Below is every line of that bill, each one the vendor's own published price, checked in September 2026 and linked so you can check it yourself.

Before you deploy anything

On your laptop: nothing

The first run needs no accounts, no keys and no card.

Postgres runs in a container on your machine. Magic links are printed to the terminal instead of being emailed, so sign-in works with no mail provider. Chat and agents fall back to Workers AI, which is the one part that calls out, and it is the only reason the first start asks you to log in to a free Cloudflare account. Start the bootstrap with --offline and even that goes away.

You can build your whole first feature, invite nobody, and pay nothing.

Once it is deployed

Four bills, and three of them are usually zero

The only one you are certain to pay is Cloudflare, and you are paying it for the plan rather than for your traffic.

Cloudflare Workers

runs the app itself

Free
Not an option here. The kit needs the Paid plan, because Workflows and the per-step CPU limits the agents run under need it.
Paid
$5 a month minimum, which includes 10 million requests and 30 million CPU milliseconds. Past that it is $0.30 per further million requests and $0.02 per further million CPU milliseconds.
Their pricing page

Postgres

every row your app owns

Free
Neon’s free plan gives you 100 compute hours a month and 0.5 GB of storage. A small internal tool fits inside that.
Paid
The Launch plan is usage based with no minimum: $0.106 per compute hour and $0.35 per GB-month of storage. The database scales to zero when nobody is using it, so an app that is quiet overnight is not billed for overnight.
Their pricing page

Email

magic links, invitations, notifications

Free
Resend sends 3,000 emails a month, capped at 100 a day, across 3 domains. Invitations and sign-in links for a team do not come close.
Paid
$20 a month for 50,000. You only need this when you are emailing customers rather than colleagues.
Their pricing page

AI

chat, agents, document search

Free
10,000 neurons a day on Workers AI at no charge. That is enough to try everything and to run a quiet internal tool.
Paid
$0.011 per 1,000 neurons beyond the daily allowance. The default chat model costs $0.060 per million input tokens and $0.400 per million output tokens; embeddings for document search are $0.012 per million input tokens. Bring your own Anthropic or OpenAI key instead and you pay them directly.
Their pricing page
A worked example

A twenty person internal tool

An estimate without its assumptions is a guess, so here are the assumptions. Twenty people who open the app a few times a day. A few thousand rows. Invitations and the occasional notification by email. Nobody is hammering the AI.

  • Workers$5. The minimum, not the usage. Twenty people will not trouble ten million requests.
  • Postgres$0. Quiet overnight and at weekends, so it sits inside the free plan's compute hours.
  • Email$0. Twenty invitations and some notifications is nowhere near 3,000 a month.
  • AI$0 to a few dollars, depending entirely on whether anyone actually uses chat.

About $5 a month, and it stays there until something changes shape.

The honest part

What actually makes it go up

In rough order of how likely each one is to surprise you.

  1. AI, and it is not close

    A chat turn is cheap. An agent that reads four documents to answer one question is not, because every passage it reads is input tokens. The kit writes a row per model call with the tokens and an estimated cost, so this is a number you can look at rather than guess at.

  2. Database compute, if the app is never idle

    Compute hours are the main Postgres cost and scale-to-zero is what keeps them low. An app with users in one timezone is cheap. A background job that polls every minute keeps the database awake all night, so do not write one.

  3. File storage, once people upload in earnest

    Uploads go to R2 and stay there. Storage is billed by the gigabyte and there are no egress fees, which is the reason R2 is used here at all, but a portal where every customer uploads their year of invoices is a real line on the bill.

  4. Requests, eventually

    Ten million requests a month is a lot for a B2B app. If you get past it you have a different problem, and it is a nice one.

Try it before you pay anything

The local run is the whole app with demo data in it. If it is not what you want, you have spent two minutes and no money.

Get it running Or weigh it against building your own