Give Ask Steve Infinite Capabilities with MCP Servers

You down with MCP?

We've been building direct integrations between Ask Steve and a number of services (Sheets, Calendar, Excel, Slack, etc.). We thought it would be cool if Ask Steve could also connect to remote MCP servers, which would unlock infinitely more integrations.

If you haven't heard of MCP, it's basically a middleware layer between an LLM and the existing API of an online service - think Gmail, Sheets, Asana, Salesforce etc. Anything that exposes an API can be made available for an LLM to use via MCP. So now aside from just processing words, the model can actually INTERACT WITH AND TAKE ACTION in other services.

Benefits of MCP

The biggest benefit of this is that you can now interact with any MCP-enabled service, versus just the ones that we have explicitly built support for. Right now we've built ~15 connections into Ask Steve, but now you can connect to THOUSANDS of applications. Zapier alone says they support 7,000+.

The biggest difference is that in the dedicated integrations we build, we are deciding ahead of time the action to be taken (e.g. "Add a Contact") and the sequence of steps to accomplish that, and using an LLM as one step in a sequence. Typically that step is only to take unstructured data and turn it into structured data that can be written into the target system.

With MCPs, you're basically telling the LLM to do something, and it will look at the MCP tools that is has at its disposal, decide if it can use any of them, and then tries to do things. You are trusting the model with the whole task, vs. just a small piece of it.

Which leads to the second big benefit - which is that you can ask for any "combination" of things and it will try to figure out how to accomplish them. This means that you don't need to know what you want to do ahead of time. For example "If the contact is named Mike and works as a recruiter, add them to Crelate, otherwise add them to Salesforce". You could never write code to support the infinite variations, but the LLM can figure it out!

Whether it works or not is another matter, but I have been consistently surprised at (a) how good the models can do sometimes (b) how bad the models can do sometimes :) You never know until you try things!

Ask Steve and MCP

After doing the research, we came to the conclusion that I shouldn't build anything, because the model decides what MCP tools to call, so we should just wait until the model providers MCP-enable their API offerings.

Which as of the last few days is now the case!

• Anthropic has a beta MCP connector for their API. You add the specs for your MCP servers to your API call and Claude does everything else behind the scenes - it connects to the servers, retrieves tools, executes the calls, keeps making calls until it's done, etc. This is nice because your client doesn't really need to do anything.

• OpenAI has added something similar to their Responses API.

• Google has also implemented something for Gemini, but their implementation seems to put more work on the client, so isn't supported by Ask Steve.

All that said, while it's now possible to CONNECT to MCP servers, doing anything that requires authentication (which is the interesting stuff) still requires work.

Let's say you want an LLM to be able to interact with a Google Sheet of yours:

• First you need to find a trusted MCP server for Google Sheets somewhere and login/create an account with that service.
• Then you'll need to OAuth authenticate with Google Sheets on that MCP server to give it access to your Google Sheets.
• Then the MCP server provider will either give you your own unique URL to use (but you better make sure no one else sees it, because if they do, they now have access to all your Google Sheets) or a generic URL and a unique access token.
• Then you need to add the url of the MCP server and optionally the access token to your API request.

For the flexibility it provides, it might be worth it! So now you know - Ask Steve can now call LLMs that support MCP to interact with other services. We added new MCP model templates for Anthropic and OpenAI that you can use to get started quickly. Watch the video above to see how to set everything up!

Questions or Problems?

Ask for help in our User Community