itsbydavid
    AI agents & MCP

    Let AI query your business directly.

    Your data lives in a database an AI cannot reach. I build the layer that changes that — a clean API, an MCP server, and agents that can actually do something with it.

    • Shipped one in production
    • Fixed scope
    • Published pricing
    • One person
    Book a 20-min call

    The problem

    An AI can only use what it can reach.

    Most business data is one step away from being useful to a model, and that step is architectural.

    Your data is locked in prose

    Prices, availability, coverage and inventory live in page copy, PDFs and images. A model can read a sentence about your pricing. It cannot filter, compare or compute with one.

    Every integration is bespoke

    Wiring your data into each assistant separately means rebuilding the same connector repeatedly. One standard interface replaces all of them.

    Agents that answer but cannot act

    A chatbot that describes what it would do is a demo. Booking, updating a record or filing the ticket is where the value is, and it needs real permissions and guardrails.

    What I actually do

    What I build

    Agent-readable APIs

    A clean, documented, read-optimised interface over the data you already have — the foundation everything else sits on.

    MCP servers

    The emerging standard for exposing tools and data to AI clients. Build it once; Claude, Cursor, Codex and anything else MCP-compatible can use it.

    Retrieval over your knowledge base

    So answers come from your documents, with citations, instead of a model guessing plausibly.

    Agents that take real actions

    Scoped permissions, audit trails and human checkpoints where the action is consequential.

    Evaluation and guardrails

    The unglamorous part that decides whether a prototype survives contact with real users. This is what I do at TELUS by day.

    The data layer underneath

    Often the real work: turning scattered, inconsistent sources into one structured store worth querying.

    Proof

    I have already shipped one.

    Parcela Nova scores every municipality in the country in Guatemala from 12+ official sources, 10+ global datasets and 5 purpose-built tools — and exposes the result through an open, keyless, read-only MCP server with eight tools any AI client can call.

    Point Claude or Cursor at it and ask about buildable land, hazard exposure or drive time to the capital, and it answers from verified data rather than guessing. Before it existed, that information lived in public registries, municipal offices, spreadsheets and brokers' memory.

    Read how it was built

    Pricing

    What it costs

    Published rates. No discovery call required to find out what things cost.

    Single System Build

    $7,900–$14,000

    61–108 hours

    4–6 weeks

    One workflow, agent, website or visibility fix, built and live.

    Your data is already structured

    $14,000–$22,000

    108–169 hours

    6–10 weeks

    Structured data layer, agent-readable API, MCP server, and the site on top.

    Your data is scattered

    $26,000–$45,000

    200–346 hours

    12–20 weeks

    The same layer, plus getting your sources into one shape before any of it can be modelled.

    See the full rate card

    Questions

    Common questions

    What is MCP, in one sentence?

    A standard way to expose your data and tools to AI clients, so you build one interface instead of a separate integration for every assistant. You are buying the outcome — your business becomes queryable — and MCP is the mechanism.

    Do we need an MCP server, or just an API?

    Often just a good API, and I will say so. MCP earns its place when you want multiple AI clients calling your data directly without bespoke glue for each one. It is a means, not the point.

    Is our data safe if an AI can query it?

    You decide what is exposed. A read-only server over a curated subset is very different from opening your production database. Scope, permissions and audit trails are part of the build, not an afterthought.

    What if our data is a mess?

    That is usually the actual project. Structuring scattered, inconsistent sources into something worth querying is most of the work — it was most of the work on Parcela Nova too.

    Could an AI answer questions about your business?

    Tell me what data you hold and what you'd want an agent to do with it. I'll tell you whether it's a weekend or a quarter.

    Book a 20-min call