Custom LLM Connectors & MCP Tools

Give a model safe access to your private data and tools.

What it is

A connector sits between a model and something you own: a knowledge base, a database, an internal API. GTDataWorks builds it as a Model Context Protocol server or an OpenAI function-calling tool, so Claude, ChatGPT or any MCP-capable client can ask your systems questions while the data stays where it is, instead of being copied into a vendor's index.

The defaults are conservative. Access is read-only unless a write is specified and gated. Each connector ships with an access policy scoped by OAuth 2.0, and every call is logged, so you can see what the model asked for, on whose behalf, and what came back.

What you get

  • Connector or MCP server with testsTypeScript or Python source, an automated test suite, and a run against a real MCP client before handover.
  • Access policyA written statement of which sources, fields and operations the model may reach, enforced in code and readable by your administrators.
  • Usage logA structured record of every call: tool, caller, arguments and time. It is written wherever you run the connector.
  • Install guideConfiguration, keys, OAuth scopes and client setup for Claude, ChatGPT or your own MCP client, written for the person who will maintain it.

How the work runs

  1. Data inventory and access policy

    You and GTDataWorks list the sources the model should reach, what is off limits, and who may ask. That list becomes the written access policy the connector enforces.

  2. Read-only adapter

    A thin adapter over your database, search index or API that can read and nothing else. Writes are added only when you ask for them, and then each one is gated and logged.

  3. MCP server or tool schema

    The adapter is wrapped as an MCP server in TypeScript or Python, or as an OpenAI function-calling tool, with tool names, descriptions and argument schemas the model can follow.

  4. Tests with an MCP client

    The server is run from a real MCP client against your data. Tests cover the expected questions, requests the policy must refuse, and malformed input.

  5. Install guide and handover

    A guide for your team covering configuration, keys, scopes and how to read the usage log, delivered with the source.

Pricing

Per connector

$1,000–$3,000

Connector or MCP server, access policy, tests and install guide.

Prices are in USD and are starting points. The final quote follows a scoping call, where the number of sources and whether any writes are needed set the figure.

Where it fits

A good fit when

  • Your internal data is valuable and must stay where it is; you want a model to query it, not a copy of it.
  • Your team already uses Claude, ChatGPT or another MCP-capable client and wants it to see internal systems.
  • You need a record of which user asked what, and what the model was allowed to see.

Not a fit when

Questions

Is the connector read-only?

By default, yes. The adapter exposes reads only. If a workflow needs a write, such as creating a ticket, it is added as a separate tool that is gated and logged, so the default stays safe.

How are OAuth scopes handled?

The connector asks for the narrowest scopes the access policy allows, and the scopes are listed in the install guide so your administrators can review them before anyone connects a client.

What is logged, and who can see it?

Every call: which tool was used, by whom, with what arguments, and when. The log is written wherever you run the connector; it is not sent to GTDataWorks.

Which clients does it work with?

Any MCP-capable client, including Claude and ChatGPT. Where MCP is not an option, the same adapter can be exposed as an OpenAI function-calling tool.

Connect a model to your data

Describe the source, who should be allowed to ask it questions, and the client your team uses. GTDataWorks comes back with a scope and a quote.