Custom MCP connectors: bring any data source to your agent

Connect your Fabrile agent to any MCP-compatible server — internal APIs, vector stores, databases, or third-party tools — without writing platform-specific glue.

•By Fabrile Team•4 min read
MCPCustom ConnectorsIntegrationsAnnouncement

Fabrile now supports custom MCP connectors. If you've built (or use) a server that speaks the Model Context Protocol, you can plug it into any Fabrile agent in a few minutes — no Fabrile-specific SDK, no custom backend work.

What this unlocks

Until now, an agent's knowledge had to come through one of our built-in sources: scraped web content, uploaded files, Google Drive, Notion, or one of the integrations on the catalog. That covers most cases, but not all. Custom MCP connectors fill the rest.

Examples we've seen during the beta:

  • Internal product search. A SaaS company connected their existing search service so the agent can look up customer-specific data instead of relying on generic docs.
  • Live inventory and pricing. An e-commerce team wired the agent to their pricing API so quotes are always current.
  • Ticketing and CRM lookups. An agent that recognizes a returning customer and pulls their open tickets from the support tool of choice.
  • Private vector stores. Teams who already invested in their own embeddings pipeline can keep it and just expose the search interface over MCP.

Why MCP

We chose MCP because it's an open standard, not a Fabrile invention. That means:

  • Any MCP server you build for Fabrile also works in Claude Desktop, Cursor, and a growing ecosystem of clients.
  • You don't have to learn yet another integration framework.
  • If you switch chatbot platforms in the future, your connector still works.

It's the opposite of a lock-in.

How to add one

  1. Open Dashboard → Integrations → Custom Connectors → New connector.
  2. Paste the MCP server URL and any auth headers it needs.
  3. Fabrile introspects the available tools and shows them in the UI.
  4. Pick which tools your agent should be allowed to call. Save.

The agent now has those tools available alongside its knowledge base. Tool calls happen server-side, on demand, with each request scoped to the connector's auth.

What about latency?

MCP tool calls add a round trip, so they're best used for things that need to be live (inventory, accounts, ticket status) rather than things that could just be indexed once (static documentation). For the latter, a knowledge base is still the right answer — it's cheaper and faster.

A good rule of thumb: if the answer can change in the next 24 hours, reach for a connector. If it doesn't, index it.

Build your own

The MCP spec is short and the official SDKs (Python, TypeScript, Go) are good. If you can stand up an HTTP server, you can stand up an MCP server. We're collecting examples on the integrations docs page if you want a starting template.