Perplexity Is Learning What I Learned Six Months Ago

At Ask 2026 last week, Perplexity’s CTO Denis Yarats said something that made me put down my coffee: his team is moving away from MCP internally because tool descriptions eat 40-50% of available context windows and authentication adds friction in production.

That’s a real number. Half your context window, gone before your agent does anything useful.

Here’s the thing — I’m not a CTO. I run a personal AI setup out of a MacBook. And I made this exact call months ago, not out of cleverness but out of watching my Claude sessions get sluggish when I had a dozen MCP servers loaded. In my setup, Kit.com, Pickaxe, MoneyWiz, and Apple Reminders all go through CLI tools now, not MCP servers. I literally keep a preference table in my Claude Code config listing which tools to route through which interface, and why.

MCP isn’t bad. It’s genuinely good for dynamic tool discovery — when your agent needs to figure out what’s available at runtime. But for tools you’re calling hundreds of times a week? You already know what they do. You don’t need to spend 500 tokens re-explaining them to the model every single session.

Perplexity is a company with real engineering resources, and they needed a conference talk to arrive at “use a CLI for predictable, high-frequency calls.” The practitioners got there first — they just didn’t write it up.

Source: awesomeagents.ai