An API is a door that code opens on exact instructions, and MCP is a standard that shows an agent which doors exist and lets it choose for itself. So check who sets the next step: if it is known in advance, go with an API, and if the agent has to decide mid-task, go with MCP. Connectors and Composio answer a different question: who holds the login to your accounts.
A small script of ours reads the open workshop dates every 30 minutes, with no model and zero tokens, because there is nothing to decide there. "Find the email from the client and book them a meeting" is a different story, since nobody knows in advance what step three will be. That freedom has a price, though: plenty of programs load their tool definitions into context before you have even said hello. In November 2025 Anthropic published an illustrative example where an agent that loads only the definitions it needs for the task drops from 150,000 tokens to 2,000 (one example, not an average, and still).
A Claude connector is a ready-made MCP connection: a few clicks, a sign-in, and Claude stays connected to the service for you. If you work with Claude yourself, start here (the docs say it is available in Claude Code when you sign in with your Claude account, so I would not build on it for a server running on an API key). It has a price too: some connectors come from outside providers, so read what you are approving.
Composio is an outside service that holds the logins to a pile of apps in one place, and agents reach them through a command line or code, from a second computer too, and with nobody at the screen. The price: one company holds access to many of your accounts.
Our agents also run at night, so on October 28 it became our first route, after a Google permission was revoked and the calendar agent told Tom twice that the calendar was down, while Composio was working as usual (the second time the agent was simply quoting itself, and yes, agents do that too). Today that holds for most of our apps, and almost every one has a backup route, because Composio gets stuck too: 80 Google Sheets tabs at once brought 4 rate-limit refusals on a shared quota, and the job finished on the backup route with nobody touching it.
The order I recommend if you work alone: a connector → Composio when agents start running without you or the apps pile up → a direct API for anything that repeats with no judgment involved. I am not sure this holds for every system, but it is where I would start today.
A prompt, on the house
I want to connect my agent to: [app name], so it can: [what it needs to do there].
Before we connect anything, help me pick one route:
1. Ask me: does this action repeat the same way each time, or does the agent need to decide along the way?
2. Ask me: am I working in Claude when this happens, or does the agent run alone?
3. Search the web for what exists today for this app: a Claude connector, an MCP server, an API, a connection through Composio. Do not trust your memory.
4. Recommend one route, with a one-line reason. Say who holds the login to my account, what it costs, and what breaks if they go down.
5. Give me a first read-only test before the agent gets permission to write or send anything.
