Intel / post

Rhys's MCP Retrospective

Source: x.com
Author: Rhys
Date: Indexed Jul 2, 2026

Context

A large tool catalog need not mean loading every tool's instructions into every AI request. In an MCP retrospective, Rhys argues that tool bloat is mainly a problem of poorly built clients, since most now load tools lazily and let the model discover them when needed. , the Model Context Protocol, is a common standard for connecting AI agents to outside tools and services.

Rhys draws on the command line as evidence. Once agents had a shell tool, they could chain commands and filter output, and the GitHub command-line tool alone exposes thousands of operations. Rhys concludes that software should accept MCP servers, command-line tools and direct APIs alike instead of forcing a choice, a premise behind the company Rhys started, Executor.

Key takeaways · AI-distilled

The 'fewer tools means better models' rule is wrong on its own terms. The GitHub CLI exposes thousands of effective actions, and effectively infinite ones once an agent uses `gh api` to hit any endpoint directly.
Friction decided this, not capability. You can tell an agent to install a CLI and use it in the same breath, while most MCP clients require a full restart before a newly added server can be called at all.
Vercel makes the case by accident: its MCP server launched with ~7 tools, sits at ~20 mostly read-only, and admits only ~16 approved clients, while its CLI runs unrestricted inside agents skipping permission prompts.
Conversion runs one way. An MCP server turns into a CLI mechanically; a CLI won't turn back, because CLIs carry no machine-readable marking of which actions are destructive the way MCP and OpenAPI specs do.
CLIs smuggle in statefulness. Most are API wrappers that got a good auth flow and no API auth flow, so an agent that only wants to read email ends up spinning a container to run the Workspace CLI.

Terms in this piece · Glossary

MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Key quotes

“there was a realization the less tools you have, the better the model does - although this isn't actually true.”

“bash benefitted from 37 years of incredible development on amazing primitives, mcp implementations were mostly vibe coded in a month.”

“you can trivially convert from an MCP to a CLI, you cannot convert from a CLI to an MCP in that same way.”

“most MCP servers could just be direct calls to an API, the same way most CLIs could just be direct calls to an API”

Read the source x.com