Perplexity Drops MCP Internally, Shifts to APIs and CLIs for AI Agents

Perplexity CTO and co-founder Denis Yarats revealed at the Ask 2026 conference on March 11 that the company has stopped using Anthropic's Model Context Protocol (MCP) internally, opting instead for traditional APIs and command-line interfaces to power its AI agents.

Why Perplexity Is Moving Away from MCP

Yarats identified two core problems with MCP that drove the decision:

• Token overhead: MCP tool definitions, including parameter schemas and response formats, consume context window tokens. For agents making many tool calls across long conversations, this overhead compounds and reduces the model's effective working memory.
• Authentication friction: Each MCP server handles its own auth flow, creating integration headaches when connecting to multiple services simultaneously.
• Schema abstraction loss: Wrapping existing APIs in MCP's standardized schema can cause information loss. Features like date ranges or snippet controls may require explicit mapping that direct API calls handle transparently.

"Every abstraction adds overhead," is the underlying philosophy. For high-throughput AI agents making thousands of tool calls, protocol translation latency becomes a real concern.

The Alternative: Perplexity Agent API

Alongside the announcement, Perplexity highlighted its Agent API, generally available since February 2026. The API offers a single endpoint that routes requests to models from OpenAI, Anthropic, Google, xAI, and NVIDIA, all accessible under one API key using OpenAI-compatible syntax.

Supported models include GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro, Grok 4.1 Fast, Nemotron 3 Super, and Sonar.

Built-in tools come with transparent pricing:

• Web search at $0.005 per call
• URL fetch at $0.0005 per call
• Function calling at no additional cost

Industry Reactions

The announcement sparked intense debate in the developer community. Some see it as validation that MCP was over-engineered from the start. Y Combinator CEO Garry Tan reportedly built a CLI instead of using MCP, citing "reliability and speed."

Others argue that MCP still serves a valuable role for smaller teams and standardized tool discovery, and that Perplexity's decision reflects pragmatic engineering trade-offs rather than fundamental flaws in the protocol.

Notably, Perplexity still maintains its official MCP server for developers who want it. The shift is internal, not a blanket recommendation against MCP.

What This Means for the AI Agent Ecosystem

The move highlights an emerging divide in AI agent architecture:

• OpenAI is betting on agent SDKs
• Anthropic continues to invest heavily in the MCP ecosystem, which has surpassed 79,000 GitHub stars
• Perplexity is championing API-first agents with minimal abstraction layers

For developers building AI agents, the takeaway is nuanced. MCP excels at standardized tool discovery and multi-agent interoperability. Direct APIs offer more control, lower overhead, and better performance for known, predictable tool integrations.

The best approach likely depends on the use case: MCP for broad ecosystem connectivity, APIs and CLIs for high-performance, tightly controlled agent systems.

Source: Awesome Agents