Free · 60 s Free AI-readiness audit See your site the way ChatGPT does.
Personal reply within one business day.

Chosen · Tools & portals

Custom AI Tools, Portals & Agent Endpoints

Calculators, configurators, quote tools and client portals that let buyers work out their own answer, with numbers you can check. AI agents can use them too, within limits you set.

from €8,900$9,800 3–8 weeks

Typical project €9k–25k.Typical project $10k–27.5k. Fixed scope, agreed in writing before work starts.

Proof on file

OWN PRODUCT · LIVE ↗

~1,100

automated tests

TotalTax, my own live calculator. An update can’t go live if any figure lacks a dated, published source.

observed 2026-10-05 · test suite count

03 Fit

Who this is for, and who it isn't.

Built for you if

  • You sell on numbers: prices, savings, tax, returns, capacity, lead times.
  • Your product has options buyers struggle to picture or combine.
  • Clients email you all week for files, documents and status updates.
  • You expect AI agents to research, quote or book for your clients, and want them to get it right.

Not for you if

  • You need a contact form or a booking widget. That’s part of a landing page.
  • There’s no real logic behind it. A calculator that always ends in “call us” annoys buyers.
  • You want a full software platform with billing and customer accounts from day one. Possible (Filyhost is one), but scoped as its own project.
  • The rules change daily and nobody owns them. A tool is only as current as its sources.

04 The problem

Your spreadsheet is your best salesperson. Nobody outside the company can use it.

Most B2B firms have one: the pricing sheet, the savings model, the options only one engineer understands. Prospects can’t use it, so they email, wait for a quote, and often buy from whoever answered first.

A public tool changes the order. Buyers get their own number in a minute, share the link with their finance director, and arrive at your call already qualified.

The hard part isn’t the screen. It’s logic that is right and tested, so a number on your website never becomes a liability.

Signs you need this

  • Every quote takes a day of back-and-forth.
  • Your pricing is too complex for a price list, so you don’t publish one.
  • Your sales engineers spend time on prospects who were never a fit.
  • Competitors have a calculator. Yours is a PDF.
  • You don’t know what an AI agent says when asked what you charge.

Before a tool: can AI tools read the site it will live on?Free AI-readiness audit · about 60 seconds.

05 What I build

Four kinds of tools, each one already running.

Calculators & quote tools

Turn your pricing into a number buyers trust

Savings, return, tax, capacity or quote calculators, every rule sourced and tested.

OWN PRODUCT · LIVE ↗ TotalTax: 15 countries, every figure sourced

Configurators

Let buyers build it before they buy it

Product builders that check what fits, draw the result and send it to sales.

OWN PRODUCT · LIVE ↗ SitShape: 155 sofa shapes in 8 families

06 What you get

What you get, item by item.

Every build includes the logic, the tests, and a way for buyers and agents to use it.

  • The tool itself

    Calculator, configurator, quote tool or client portal, built for phones first and fast everywhere.

    Helps with: Chosen by buyers
  • Numbers you can check

    Every rule traced to a source. Automatic tests stop an update if a number breaks or a source goes missing.

    Helps with: Runs for you
  • Results to share

    Each result has its own link that opens exactly the same. Optional: results by email, or a quote request with every input attached.

    Helps with: Chosen by buyers
  • Payments

    Deposits, licences or subscriptions by card, where the tool sells directly.

    Helps with: Chosen by buyers
  • Access for AI agents (optional)

    Agents can look things up, quote or book within limits you set; anything binding waits for a person.

    Helps with: Found by AI
  • Pages that explain the method

    How the tool calculates, with sources, so AI assistants can quote the method too.

    Helps with: Quoted by AI
  • Security and privacy

    Limits against abuse, bot protection, a locked admin area, and nothing stored that you don’t need.

    Helps with: Everything

07 How it runs

Logic first, screens second.

The calculation is built and tested on your real examples before any design. If the numbers are wrong, nothing else matters.

Agents at work Human gate: nothing moves on until a person approves

  1. Step 1: Week 1

    Spec

    Inputs, rules, sources and unusual cases written down, plus a clickable outline of the screens.

    • agent://researcher
    • agent://architect

    Human gate You approve the spec and the source for every rule.

  2. Step 2: Weeks 2–3

    The calculation

    Built and tested on its own, with test cases from your real quotes and examples.

    • agent://builder
    • agent://tester

    Human gate You check the numbers against cases you know by heart.

  3. Step 3: Weeks 3–5

    Screens

    Phone-first screens, result pages, sharing and enquiry capture, on a preview link.

    • agent://builder

    Human gate You approve the screens on the preview.

  4. Step 4: Weeks 5–7

    Agents and connections

    Optional AI-agent access, payments, CRM or email. Security and load checks.

    • agent://builder
    • agent://tester
  5. Step 5: Weeks 6–8

    Launch and handover

    Live on your domain and monitored. A written guide, and your team shown how to update the sources.

    • agent://watcher

    Human gate You sign off go-live.

08 Proof

Tools I run myself, live and public.

Real products, not mock-ups. Open them and try the numbers.

9✓ Vowko: actions members’ AI agents can take observed 2026-10-05 · codebase
8✓ Filyhost: limits that stop abuse observed 2026-10-05 · wrangler config
0.1%✓ Oil Fund Ledger: max gap to the fund's reported value over 28 years observed 2026-10-05 · reconciliation gate

Also live: Vowko, where members’ AI agents can act but never vouch for anyone; Filyhost, a client file portal; and Oil Fund Ledger, a public data tool that won’t publish if its numbers drift from the fund’s own reports.

09 Price

What moves the price, and what it costs to run.

Starts at €8,900$9,800; most tools land at €9k–25k$10k–27.5k. A price range in writing within 48 hours of our call; the fixed price follows the one-week spec, which is part of the project.

Costs more when

  • Many rules, countries or product variants.
  • Accounts, roles and a client portal.
  • Payments and order handling.
  • Access for AI agents.
  • Connections to your CRM, ERP or quoting system.

Costs less when

  • Logic already working in a spreadsheet you trust.
  • A public tool with no accounts.
  • One language and one currency.
  • One person on your side who owns the rules.

Monthly running costs

Paid by you, directly to each provider, in your own accounts. I add no markup and no hosting fee.

WhatHow it worksPaid to
HostingCloudflare. Small volumes fit the free plan; the paid plan is budgeted when use grows.Cloudflare
PaymentsStripe’s standard fees, only if the tool sells.Stripe
AI model usageOnly if the tool uses an AI model (to explain a result or draft a quote), capped by a monthly budget.Anthropic / OpenAI

10 Handover

What you own when we're done.

  • Source code and tests your GitHub organisation
  • The tool, its data and access keys your Cloudflare account
  • The spec, with every rule and its source with the code
  • Domain, analytics and admin access your accounts
  • Guide to updating rules and sources with the code

No licence fees to me, no lock-in, no proprietary platform. If I disappeared tomorrow, any competent developer could pick it up from the written instructions.

11 Under the hood

For your web developer, the technical detail.

Show the technical detail
Stack
Cloudflare Workers, D1 for data and R2 for files where needed (Workers Paid when volume needs it). The calculation engine is a pure TypeScript module, separate from the interface.
Rules and tests
Rules live in data, each figure with its source and date. The test suite runs on every change and fails the build if a number breaks or a source is missing.
Results and method
Inputs in the URL, so a result link reopens exactly; crawlable HTML method pages for each calculation and its sources.
Agent access
An MCP server or WebMCP tools: scoped, logged, rate-limited actions with a spending cap; anything binding is human-only.
Security
Edge rate limits, bot protection, Cloudflare Access in front of admin, minimal data stored.

12 Questions

Straight answers to fair questions.

Does a calculator actually need AI?

Usually not for the calculation itself: that should follow fixed rules and be tested. AI helps around it, explaining a result in plain words or drafting the follow-up quote.

Can AI agents use the tool, and is that safe?

Yes, through MCP, a standard way for AI assistants to use a service, and only within limits you set: a short list of actions, caps on how often and how much, and a log of everything. Anything binding can be marked “person only”, the way Vowko lets agents act but never vouch. Not needed yet? It can be added later.

How do you make sure the numbers are right?

Every rule is written down with its source before it’s built, then tested against your own examples. On TotalTax, ~1,100 automatic tests run on every change, and an update can’t go live if a figure lacks a dated source.

Who maintains it when the rules change?

Your team or me. Rules live in data, so changing a rate or a price is an edit, not a rebuild: your team does it with the written guide, or I do through the AI Department retainer.

Rather talk it through first?

20 minutes with me, not a sales rep. You leave with a clear next step — whether or not we work together.