Making MCP cheaper via CLI

Original Article ↗ Hacker News Discussion ↗

CLI vs MCP: effectiveness and context cost

  • Many commenters report better real-world performance with CLIs than with MCP tools, especially for coding agents where a human reviews actions.
  • Core complaint about MCP: every tool’s JSON schema bloats the context window on each request, especially when dozens of tools are loaded.
  • With CLIs, the model only needs to discover commands via --help and can then call them directly; tool descriptions don’t sit in context every turn.
  • Some argue the article understates improvements like Anthropic’s tool search, which avoids dumping all MCP definitions, but others note most existing MCP servers still behave in the “dump everything” style.

Composability, behavior, and training

  • CLIs shine because models are heavily trained on shell usage: piping, grep, jq, loops, etc.
  • This enables powerful single-call workflows (e.g., looping over many IDs and aggregating results) that would otherwise require many separate MCP tool calls and blow up context.
  • CLI output is easily filtered (head, jq, etc.), reducing token usage versus large JSON dumps from MCP tools.

Arguments in favor of MCP

  • MCP provides a standard protocol for auth (notably OAuth with dynamic client registration) and session management, which matters for third‑party SaaS and consumer products.
  • It lets upstream providers expose problem-oriented tools (e.g., “get meeting transcript” instead of multiple low-level API calls), improving multi-step workflows.
  • MCP is the only supported integration path for some general assistants (e.g., ChatGPT/Claude) that cannot run arbitrary CLIs.

CLI, Skills, and hybrid approaches

  • Many prefer a hybrid: MCP servers running system‑wide, with thin CLIs (or shims like mcpshim, CMCP, mcp-cli, MCPorter, CLIHub) exposing them as shell commands.
  • Skills/AGENTS.md/markdown “skill files” describing available CLIs are seen as a lightweight alternative to MCP for tool discovery and progressive disclosure.
  • Existing vendor CLIs (GitHub, AWS, Atlassian, etc.) are often considered superior to their official MCPs; some question why to wrap MCPs at all instead of using those directly.

Architectural and future considerations

  • Biggest cost driver is repeatedly resending long conversation histories, not just MCP schemas; batching and parallel tool calls help.
  • Several commenters think the real long-term fix is better attention mechanisms, larger cheap context windows, and persistent, cached system prompts—rather than protocol swaps.