Your MCP Server Is Eating Your Context Window. There's a Simpler Way
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
MCP tool definitions can consume 55,000+ tokens before an agent processes a single user message, with benchmarks showing MCP costing 4–32× more tokens than CLI for identical operations. A CLI-based approach using progressive disclosure (--help flags) reduces the agent system prompt to ~80 tokens, with capabilities discovered on demand at 50–200 tokens per lookup. The Apideck CLI demonstrates this pattern: a static Go binary that parses OpenAPI specs dynamically, auto-detects TTY vs non-TTY for output format, and enforces structural safety (GET auto-approved, POST prompts, DELETE blocked) in code rather than prompts. The post also covers when MCP still wins: tightly scoped high-frequency tools, complex stateful workflows, and multi-tenant B2B scenarios requiring per-user OAuth and audit trails.
Table of contents
The problem nobody talks about at demo scaleThree approaches to the same problemWhy CLIs are the pragmatic sweet spotHow we built itWhen CLI isn't the answerWhat this means for API providersFurther reading1 Comment
Sort: