Service · AI integrations

A custom MCP server, so AI knows your system.

MCP (Model Context Protocol) is the open standard AI assistants use to talk to other software. We build an MCP server for your API, CRM, store or database, so your team can ask Claude, ChatGPT or Cursor about real company data and hand it tasks, with permissions and a log of every call.

What problem does it solve?

Your team already uses Claude or ChatGPT, but the assistant knows nothing about your business. To answer “which customers have not ordered in three months?”, someone exports a CSV, pastes it into a chat and hopes nothing sensitive went along with it.

An MCP server flips that around. The assistant fetches what it needs through a controlled connection, only within the scope you allow, and every request is logged. One server works with every MCP-compatible tool, so you do not build a separate integration for each assistant.

How it works, step by step

  1. Agree the scope

    Which questions the team wants to ask and which actions to delegate. Most projects start read-only: search customers, orders and documents, pull reports.

  2. Design the tools

    Each tool is one clear operation with a description and input validation, such as “find a customer’s orders” or “create a draft quote”. The model knows when and how to use it.

  3. Permissions and sign-in

    OAuth or per-user API keys, scopes that mirror your own system, and any write action only after a human confirms it.

  4. Logs and limits

    Every call is recorded (who, when, what), with rate limits and prompt-injection hygiene: content coming from your data is treated as data, never as instructions.

  5. Deploy

    Locally for developers (stdio) or hosted over HTTP with sign-in for the whole team, plus setup notes for Claude, ChatGPT and Cursor.

  6. Hand over and extend

    You get the code and docs. When the team finds a missing tool, we add it.

What it does not do (and when a human steps in)

  • No “just in case” full access. Only the operations the agreed tasks need.
  • No changes without confirmation: deleting, payments and anything sent to customers always need a human decision.
  • We do not store your data. The server runs on your infrastructure or account; secrets live in your secret manager.
  • No unofficial access such as scraping admin panels. Official APIs only.

What it connects to

  • Claude (desktop and claude.ai)
  • ChatGPT
  • Cursor
  • REST and GraphQL APIs
  • CRMs with an API
  • Shopify, WooCommerce
  • Databases
  • Google Workspace

We connect only through official APIs and only to what we check during scoping. Product names are used to describe integrations.

How much does it cost?

From $300 for the basic package on Fiverr and Contra. Fixed price, agreed scope, no hourly surprises.

PackagePriceWhat you get
Basic$300Up to 3 tools, read-only, runs locally (stdio), README.
Standard$750Up to 8 tools including write actions with confirmation, hosted HTTP server with OAuth or API key, docs.
Premium$1,400Standard plus an npm package, directory listings (e.g. Smithery), per-key spending limits and logs.

The price depends on the number of tools and on whether the server only reads data or also changes it. Prices match our Fiverr packages; on Contra the same service starts at the basic price.

Bigger or custom scope? Send us two or three sentences about the process and we will reply with a fixed quote. Your AI model and API costs are paid directly by you, so you always see them.

How we work together

  1. Scope: a short call or a few messages. We agree what the agent does, what it connects to and what “done” means.
  2. Build at a fixed price: through Fiverr and Contra, with milestones you can review. We test on your real examples before handover.
  3. Handover: you get the code, the setup guide and the keys in your own accounts. Ongoing maintenance is optional and quoted separately.

Compliance and house rules

  • GDPR: a data processing agreement, minimal data scope and logs with a limited retention period.
  • Security: least privilege, secrets outside the code, human confirmation for anything that changes data.
  • EU AI Act, Art. 4: a short training for the team on how to use the assistant and where its limits are.

What our claims are based on

We built and published our own MCP server for our app: the fattly package on npm (run it with npx fattly mcp), also available as a hosted server with OAuth and listed on Smithery. Anyone can check it.

We have not yet deployed MCP servers inside client companies, so we do not quote client results.

Frequently asked questions

What is an MCP server?

A small program that exposes your system’s data and actions to AI assistants through the open Model Context Protocol. Claude, ChatGPT or Cursor can then use your CRM, store or database in a controlled way.

Which AI tools work with an MCP server?

Tools that support MCP, including Claude (desktop app and claude.ai), ChatGPT and Cursor. One server serves all of them.

Can the assistant break something in our system?

Only within the scope you allow. We usually start read-only, and actions that change data require a human to confirm. Every call is logged.

How is MCP different from a regular API integration?

A regular integration wires two specific systems together. An MCP server exposes operations in a form an AI assistant can choose to use when it needs them, and it works with several assistants at once.

How much does a custom MCP server cost?

From $300 for a small read-only server with up to 3 tools, $750 for a hosted server with write actions and OAuth, and $1,400 with an npm package, directory listings and spending limits.

Where does the server run?

Locally on developers’ machines or hosted on your own cloud account with sign-in for the team. Your data does not pass through our servers.

Contact

Tell us what eats your team's time.

Two or three sentences are enough: what your business does and which process you would hand over first. For fixed-price packages, use Fiverr or Contra.