MCP Connectors
MCP is the open standard AI assistants use to work with other software. We build the connector that lets Claude, ChatGPT and your own agents read from your application and act in it, within the permissions you set. You don't rebuild your application.
YachtBooks, our own product, has one: ask Claude about budgets, invoices and maintenance, or log an expense, in the same conversation.
What we do
- Tools shaped by your domain
- We learn how your business works, then design tools around the jobs people do, so the AI isn't guessing its way through raw API endpoints.
- Read and act
- Answer questions from your live data and, where you allow it, create records, update information and start workflows.
- Permissions that hold
- Separate read and write access, per-user sign-in, audit logs and human approval for sensitive actions, all enforced on the server.
- Connected to what you have
- ERP, CRM, SaaS or a proprietary platform. We work through your existing APIs and business rules, then test, deploy and support the connector.
Built in our own product
YachtBooks
Budgets, expenses, approvals, invoices and planned maintenance for yachts and fleets, in one place.
Common questions
- Do I need to rebuild my application?
- No. In most cases we connect through its APIs. Where those fall short, we add a dedicated integration layer.
- Will it work with ChatGPT and Claude?
- Yes, if the platform supports the MCP features the connector uses and your plan allows custom connectors.
- Can AI change data in my application?
- Only if you decide it can. Read and write access are designed separately, tool by tool, and sensitive actions can wait for a person's approval.
- Does our data reach the AI provider?
- Yes. Whatever a tool returns goes to the assistant's model and falls under that provider's data terms. So we decide with you what each tool returns, and leave out what shouldn't leave your systems.
- Does the AI act as the user, or through a shared account?
- We prefer per-user sign-in wherever your application supports it. The AI then sees only what that person can see, and each action is logged under their name.
- What stops prompt injection from triggering an action?
- A record or document can contain text written to mislead the model, so the connector doesn't trust the model. Permissions are checked on the server for every call, and the actions that matter can require a person's approval.
Other services
- Product Development6 products of our own, built and run by the same team that builds yours.
- Post-Acquisition IntegrationLate-stage companies above $100M ARR, integrated and brought to 75–85% profit margins.
- Cloud Cost OptimizationAWS bills cut by 67% on average, and $15M saved for clients so far.
- AI & Digital TransformationAI built into three of our own products: App Runner, SyncTalk and Borealis.
Tell us what you need built, merged, or made cheaper.
Talk to usor email info@devfabrika.com