Anthropic's Universal Protocol Falters as MCP Hits 97M Downloads Yet Can't Link Two Agents

Anthropic’s Model Context Protocol (MCP) has become a de‑facto standard for connecting large language models to external tools, yet its limitations are now surfacing in production environments. According to a March 15 post by Calin Teodor, the protocol has amassed 97 million monthly SDK downloads and powers thousands of integrations—from GitHub and PostgreSQL to Slack and web search—but it still cannot link two autonomous agents. “MCP connects agents to tools. It does not connect agents to agents,” Teodor writes, highlighting a structural gap that forces developers to cobble together ad‑hoc communication layers such as HTTP endpoints, message queues, or shared databases for peer‑to‑peer interaction.

The missing horizontal layer, dubbed the “Pilot Protocol” by the same analyst, is essential for true multi‑agent collaboration. While MCP supplies the “eyes and hands” of an AI—access to databases, APIs, and file systems—the Pilot driver is supposed to give the agent a “voice and ears,” handling discovery, trust, encrypted tunnels, and task delegation. Without this, each MCP deployment must reinvent peer communication, adding operational overhead and security risk. The dual‑client architecture described by Teodor—one MCP client for tool access and one Pilot client for inter‑agent messaging—underscores how the two protocols operate independently, with separate endpoints and lifecycles, complicating any effort to upgrade tools without touching the networking layer.

Industry insiders are echoing these concerns. At Perplexity’s Ask 2026 conference, CTO Denis Yarats publicly admitted that his engineering team is moving away from MCP in favor of “plain APIs and CLIs,” the same “tools we’ve had for 30 years,” according to Mila Kowalski’s March 15 report. Yarats’ remarks echo a broader sentiment among production engineers who have been “discovering for months” that MCP’s promise of a universal “USB‑C for AI” adapter collapses under real‑world load. The “context window tax”—the overhead of transmitting entire schemas and parameter definitions for each tool—has become a practical pain point, further eroding the protocol’s appeal.

The backlash has spilled onto social media and the broader tech community. Garry Tan, president of Y Combinator, labeled MCP “sucks honestly” in a tweet the same day Yarats spoke, while entrepreneur Pieter Levels declared the protocol “dead.” The debate has turned Twitter/X into a “warzone,” as Kowalski notes, reflecting a growing disillusionment with a standard that has not delivered on its most critical promise: seamless agent‑to‑agent coordination. This sentiment arrives amid Anthropic’s legal battles, including a Reuters‑reported lawsuit to block a Pentagon blacklist that would restrict the company’s AI deployments. The litigation underscores the strategic stakes for Anthropic, which is simultaneously defending its market position and confronting technical criticism of its flagship protocol.

Analysts warn that the inability to natively link agents could hinder Anthropic’s broader ecosystem ambitions. As enterprises look to build autonomous workflows that span multiple AI services, the need for a robust horizontal communication layer becomes a prerequisite. Without it, developers must invest additional engineering resources to stitch together disparate components, diluting the cost‑and‑time advantages that MCP was marketed to provide. If the industry continues to gravitate toward lightweight, well‑understood APIs and CLIs—as suggested by Yarats—Anthropic may see its influence wane, especially as competitors double down on end‑to‑end agent frameworks that already incorporate both vertical (tool) and horizontal (peer) connectivity.