Writing

Is MCP Dead? The Context Crisis That Broke Naive Tool Loading. Agent Skills vs. MCP vs. CLI

Navigating the Challenges of MCP: From Adoption to Context Management. Did Agent Skills and CLI kill MCP?

Rick Hightower

Originally published on Medium.

Navigating the Challenges of MCP: From Adoption to Context Management. Did Agent Skills and CLI kill MCP?

AI Agents MCP context crisis: too many servers flood the context window while a selective harness keeps the working set small AI Agents MCP context crisis: too many servers flood the context window while a selective harness keeps the working set small

Is the Model Context Protocol (MCP) really dead? Dive into the architectural evolution that reshaped AI tooling, the pitfalls of naive tool loading, and discover why the future of AI agents is brighter than ever. Don’t miss this critical analysis of the context crisis that transformed the game! Read more here!

Anthropic says it had seen tool definitions consume 134K tokens before optimization (Anthropic, November 24, 2025).

  • More startup tokens burned before work begins
  • More ambiguity during tool selection
  • More chances to choose the wrong tool or wrong parameters
  • More irrelevant intermediate results entering the window
  • Less working room for the code, logs, requirements, and constraints that actually matter

The model is not just “holding” the tool list. It is reasoning over it.

AI Agents startup context comparison: eager loading burns startup tokens while tool search keeps the visible toolset narrow AI Agents startup context comparison: eager loading burns startup tokens while tool search keeps the visible toolset narrow

  • A command surface humans already trust

  • Built-in help and discoverability through --help, manuals, and examples

  • A place to filter, aggregate, or transform data before it ever reaches the model

  • when to use the CLI

  • which commands are safe

  • which outputs to filter

  • which scripts to run first

  • what success looks like

  • release procedures

  • incident response playbooks

  • PR review workflows

  • deployment checklists

  • codebase-specific build and test sequences

  • internal research routines

  • use MCP when you need shared, authenticated, standardized access to remote systems

  • use skills when you need workflow memory, behavioral guidance, or bundled scripts

  • use CLIs when the operational surface already exists and the shell is the most efficient front door

  • Traditional large-tool setup: about 77K tokens consumed before real work begins

  • Tool Search Tool setup: about 8.7K total context consumption

  • Savings: about 85%

  • Context preserved: about 95% of the window

The answer to large tool libraries is not “never use MCP.”

The answer is “stop treating discovery like a startup event.”

  • keep a small always-visible surface

  • expose the high-frequency operations directly

  • let deeper functionality load only when needed

  • standard connectivity

  • uniform authorization

  • attached documentation

  • reusable server contracts

Agentic AI hybrid harness stack: Agent Skills guide behavior, MCP handles remote systems, and CLI plus code keeps noisy data out of context Agentic AI hybrid harness stack: Agent Skills guide behavior, MCP handles remote systems, and CLI plus code keeps noisy data out of context

  • Use MCP for remote systems that genuinely need standardized auth, shared reuse, or rich resource access.

  • Put large or noisy tool surfaces behind search, toolsets, or code-execution front doors.

  • Use Agent Skills for workflow logic, domain playbooks, operational habits, and bundled scripts.

  • Reach for CLIs when the shell already gives you a mature operational interface and the task is mostly procedural.

  • Partition agents by job, not by ego. Research agent. GitHub agent. Deployment agent. Workspace admin agent. Not one giant universal assistant with everything attached.

  • Treat context as cache. Anything that can live in a file, script, summary, or execution environment should not sit in the hot path by default.

  • eager-loading entire server catalogs

  • exposing every tool to every agent

  • sending intermediate junk back through the model

  • assuming larger windows erase attention problems

  • confusing interoperability with good harness design

  • MCP for standard integration

  • Tool Search for discovery

  • Code execution for orchestration

  • Skills for behavior

  • CLIs for proven operational surfaces

  • harnesses that keep each AI agent narrow, sharp, and accountable

Strong AI agents are not the ones that see everything.

They are the ones that see the right thing at the right moment, and nothing else gets to waste the window.

#MCP #Model Context Protocol #context crisis #tool loading #Agent Skills #CLI #AI architecture