# MCP Doesn't Suck. Your Agent Does. — Jan Čurn, Apify **URL:** \ **Category:** Content Creators\ **Tags:** mcp, agents, ai-engineer\ **Created:** [2 October 2026 14:00 UTC](https://www.artofsm.art/t/mcp-doesnt-suck-your-agent-does-jan-curn-apify/24869 "2026-10-02T14:00:05Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [2 October 2026 14:00 UTC](https://www.artofsm.art/t/mcp-doesnt-suck-your-agent-does-jan-curn-apify/24869/1 "2026-10-02T14:00:05Z")
[![](https://www.artofsm.art/uploads/default/original/3X/2/b/2bdda4e7199d3804992c3d555b473a9544c08e18.jpeg "MCP Doesn't Suck. Your Agent Does. — Jan Čurn, Apify") ](https://www.youtube.com/watch?v=pAnLpiAG6Es) Jan Čurn argues that the criticisms of the MCP protocol stem from poor agent implementations rather than flaws in MCP itself, advocating for smarter tool management strategies like progressive tool discovery and treating MCP tools as code. He introduces MCPC, a universal CLI client for MCP that combines the strengths of MCP and CLI tools to improve efficiency, reduce token usage, and enhance AI agent-tool interactions. ---
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [2 October 2026 14:24 UTC](https://www.artofsm.art/t/mcp-doesnt-suck-your-agent-does-jan-curn-apify/24869/3 "2026-10-02T14:24:18Z")
In his talk titled “MCP Doesn’t Suck. Your Agent Does,” Jan Čurn, founder and CEO of Apify, addresses the mixed reputation of the MCP (Machine Control Protocol) standard, which was introduced by Anthropic nearly two years ago to facilitate secure interaction between AI agents and tools. Despite its rapid adoption and widespread use—boasting thousands of servers and becoming a standard for connecting AI agents to tools—MCP has faced significant criticism. Many prominent voices in the AI community have labeled MCP as flawed or ineffective, often blaming the protocol itself rather than the way agents implement it. Čurn explains that much of the criticism stems from how agents handle MCP, particularly the naive approach of loading all available tools into the agent’s context at once. This leads to excessive consumption of tokens, bloated contexts, reduced accuracy, and higher costs. However, he emphasizes that these issues are not inherent to MCP but rather to the agent implementations, as the MCP specification leaves harness design up to developers. To address these problems, several solutions have emerged, including splitting contexts into sub-agents, progressive tool discovery (loading tools only when needed), and treating MCP tools as code rather than simple functions, which leverages models’ strengths in code analysis. The talk contrasts MCP with CLI (Command Line Interface) tools, highlighting that agents naturally use CLI tools progressively and treat them as code, which is more efficient. CLI tools have been optimized over decades for concise and effective communication, and agents are well-trained in their use. However, CLI tools lack a standardized remote access protocol, making MCP more suitable for remote interactions. Čurn proposes a hybrid approach: using MCP for remote access and CLI for local agent interactions, combining the strengths of both. To demonstrate practical advancements, Čurn introduces MCPC, a universal CLI client for MCP developed by Apify. MCPC supports the full MCP protocol, including advanced features like asynchronous tasks and progressive tool discovery. It acts as a lightweight wrapper that abstracts MCP complexities into a CLI interface, making it accessible for both humans and agents. MCPC also supports JSON output for easy scripting and integration, and it securely manages credentials. The tool aims to bridge the gap between MCP’s potential and current agent capabilities, offering a more efficient and feature-rich client experience. Finally, Čurn shares performance benchmarks comparing MCPC, raw MCP, and CLI tools, showing that MCPC and CLI perform similarly in terms of speed and token cost, while raw MCP tends to consume more tokens despite sometimes finishing faster. He encourages the community to stop dismissing MCP outright and instead embrace a combined MCP plus CLI approach, which he argues is the most effective way forward for AI agent-tool interactions. The talk concludes with an invitation to explore and contribute to MCPC and related projects. ## Useful Links - [MCPC (GitHub)](https://github.com/apify/mcpc) — Directly enables action on the main claim by providing the tool demonstrated and discussed in the talk.