MCP Doesn't Suck. Your Agent Does. — Jan Čurn, Apify
The talk argues that recent criticism of the Messaging/Communication Protocol (MCP) is misplaced; the issue lies with naive client implementations (the 'harness') rather than the protocol itself. Jan Čurn outlines three key architectural fixes—Sub-agents, Progressive tool discovery, and Code Mode—to solve context bloat. He advocates for a hybrid approach: using MCP for secure, remote access and leveraging Command Line Interfaces (CLIs) for local agent interaction, demonstrating this with the open-source universal CLI client, mcpc.
Key takeaways
-
Context Bloat is a Harness Issue
3:52
The primary complaint—that MCP consumes too much context (e.g., loading 100 tools at once)—is due to poor client implementation, not the protocol design. The protocol itself is robust, but the agent harness must be designed correctly.
-
Three Architectural Fixes
5:46
Solutions include: 1) Using sub-agents to delegate tasks and isolate context. 2) Implementing Progressive tool discovery, where tools are added to context only when needed. 3) Adopting Code Mode, treating tools as code that models are better at analyzing and writing than traditional tool calling.
-
CLI vs. MCP: Local vs. Remote
11:56
CLIs are superior for local access because agents naturally interact with them as code (like bash), and they do so progressively. MCP remains the superior choice for secure, authenticated remote access.
-
Introducing mcpc
17:25
mcpc is an open-source universal CLI client for MCP, designed to support all protocol features (e.g., stddio, remote servers, OAuth, asynchronous tasks, and d-JSON output) while abstracting the complexity of the MCP protocol for both agents and humans.