I build MCP into my apps. My own agents almost never use it. My setup is simple. Each agent is a Claude Code session with its own project folder and a .env containing whatever keys it needs. I tell i
Why CEREBRO kept it
Practitioner note on agents ignoring MCP; tool-use pattern
The text below is an automated extraction of the article at https://x.com/stevelacey/status/2117359574181163199, stored verbatim in the public cerebro-vault repository. Copyright remains with the original publisher (x.com).
My setup is simple. Each agent is a Claude Code session with its own project folder and a .env containing whatever keys it needs. I tell i
> Practitioner note on agents ignoring MCP; tool-use pattern
I build MCP into my apps. My own agents almost never use it.
My setup is simple. Each agent is a Claude Code session with its own project folder and a .env containing whatever keys it needs. I tell it to get something done. I don’t care how.
It almost never picks MCP.
Not because MCP is bad, but because of the connector step. In most setups, a human has to install and configure the connector before the agent can use it. The agent can’t do that part, so it stops and asks me to get up, go to the computer and click things.
Meanwhile, it could just call the API or open a browser. So that’s wha
Backlinks
Appeared in 1 briefing