Everyone's Quitting MCP… Here's What They're Using Instead #shorts

Why It Matters

MCP’s token overhead threatens model performance and cost efficiency, prompting developers to adopt leaner CLI/API approaches that reshape how AI agents integrate tools.

Key Takeaways

• MCP's token overhead inflates context windows, hurting model performance.
• GitHub MCP alone consumes ~50,000 tokens for tool definitions.
• Builders favor CLI and API approaches to avoid MCP's context tax.
• Protocol itself works, but scaling issues arise with many integrated tools.
• Indexing and consolidation can mitigate MCP drawbacks, but trade‑offs remain.

Summary

The video highlights a shifting sentiment among AI developers: the once‑popular Model Control Protocol (MCP) is being abandoned in favor of leaner CLI and API‑only solutions. Recent posts from industry figures, including Y Combinator’s Gary Tan and the CTO of Propexity, underscore a growing consensus that MCP’s design creates more problems than it solves.

The core issue is token bloat. MCP requires loading full tool descriptions—names, parameters, documentation—into the model’s context before any user input. The GitHub MCP integration alone bundles over 90 tools, consuming roughly 50,000 tokens, which dramatically shrinks the window available for actual task data. As more MCPs are stacked, the cumulative “MCP tax” inflates input costs and introduces noise that degrades model performance.

Despite these drawbacks, the protocol isn’t dismissed as fundamentally flawed. It successfully unified disparate integrations under a single standard, eliminating the need for custom code per tool. However, scaling challenges persist, prompting developers to adopt indexing and consolidation techniques to trim context size. The video notes that many now prefer direct CLI calls, which sidestep the heavy token load.

For AI product teams, the trend signals a move toward more efficient agent architectures. Reducing context overhead can lower inference costs, improve response accuracy, and accelerate development cycles. As the ecosystem evolves, tooling that balances integration simplicity with token efficiency will likely dominate.

Original Description

"MCP sucks honestly." — Garry Tan, Y Combinator.

That one tweet kicked off the biggest shift in how we build AI agents in 2026. Perplexity's CTO followed up saying they're moving to APIs and CLI only. So what happened to the "USB-C for AI agents"?

Turns out, MCP loads every tool description into your context window before the user even types — GitHub's MCP alone burns 50,000 tokens. More context = more noise = worse model performance.

But is MCP really dead, or are we just bad at scaling it? Watch the full breakdown.

#MCP #CLI #AIAgents #LLMEngineering #DevTools