MCP vs CLI: why AI developers are moving away from MCP

If you use Claude Code, Cursor or any AI coding agent, you've come across MCP — the Model Context Protocol that was supposed to change everything. But in 2026, the developer community is moving away from MCP in large numbers. Once you understand why, it's hard to blame them.

The problem with MCP (it isn't pretty)

On paper, MCP was appealing: let your AI agent control Stripe, GitHub, Linear, Context7… In practice, though, it means loading ALL those tools into your agent's context at the start of the conversation. Even the ones it doesn't need.

Imagine opening your toolbox to hammer in a nail and being forced to carry a chainsaw, a blowtorch and a cement mixer too. Nice, right?

Each MCP server injects a large description into the model's context: how to use each tool, which parameters to pass and so on. Even if your project doesn't use Linear, the model carries Linear's instructions throughout the ENTIRE conversation. The result: it eats tokens, slows responses and makes the model less accurate because it's drowning in unnecessary instructions.

Anthropic's solution? Replace automatic loading with "MCP Search", a single tool that finds the right MCP when the agent needs it. But then the model no longer knows which tools are available, so it uses them… much less. Not the best fix.

And I almost forgot the best part: MCP can also mean an npm process running in the background on your machine each session. It uses RAM and CPU, and sometimes crashes without warning. If you've had an MCP server that works every other day, you know what I mean.

The solution: skills and CLI tools (a powerful combination)

The real breakthrough is combining skills and CLI tools. It's almost embarrassingly simple.

Choose skillis a small text file describing a capability. For example: "If you need technical documentation, use this CLI command." The model loads the file ONLY when it needs it, not before. That changes everything.

And the CLI (Command Line Interface) is the tool itself: a program run from the terminal. No MCP server running in the background eating RAM. No mysterious npm process. Just a reliable, fast bash command doing its job 365 days a year.

The best part? The AI agent already has native access to bash. It's one of its basic tools. Instead of adding abstraction on top of abstraction (MCP → tool → action), let the agent do what it already knows: run commands.

The figures speak for themselves: Playwright in CLI mode uses 90% fewer tokens than its MCP equivalent to do exactly the same thing. Ninety per cent. That's not a small difference.

Ten CLI tools that make a difference

Here's my selection of the most useful CLI tools for your AI agent in 2026:

CLI Anything — the meta-tool. Point it at an open-source project (Blender, OBS, NotebookLM…) and it automatically creates a CLI to control it from the terminal. Pure genius.

NotebookLM-PI — connects Claude Code to Google NotebookLM. Provide YouTube URLs, get a full analysis, create podcasts, slides and quizzes… all without spending a token, because Google does the work on its servers. An absolute treat.

Stripe CLI — if you've struggled with Stripe's interface (who hasn't?), this CLI is a blessing. No more 20 tabs to create a product.

FFmpeg — video, audio and subtitle processing. Ideal for creating web animations from videos. I use it for custom scrolling effects on client websites.

GitHub CLI — if you don't have this yet, you're behind. Pushes, commits, branches, pull requests… all from the terminal in one line.

Vercel CLI — simpler deployments with a generous free tier. GitHub CLI and Vercel CLI give you a CI/CD pipeline with two tools. Job done!

Supabase CLI — databases and authentication, all open source. Since Supabase can run locally, the CLI handles that too.

Playwright CLI — browser automation. The agent can launch its own Chrome instances to test forms and check pages without you lifting a finger. And remember: 90% fewer tokens than MCP.

LLM Fit — want to run AI locally but can't choose among Ollama's 400 options? This CLI analyses your machine and tells you which model will run best.

Google Workspace CLI (GWS) — the final boss. Emails, Docs, Sheets… your agent controls your whole Google workspace. With security guardrails (Google Armor handles prompt injection), it's seriously powerful.

Creating your own CLI tools (yes, you can)

This is where it gets REALLY interesting. With a tool such as API to CLI, you can turn any API into a CLI your agent can use. Point it at a service's Swagger documentation and it automatically generates a complete CLI with all the commands.

The huge advantage over MCP? You can customise your CLI tools. Want email signature management in your mailing CLI? Add it. Try doing that with a fixed MCP integration…

Chaining tools together works naturally too. "Research React using Context7, create a tweet with Typefully, then send a newsletter": one conversation, each skill loaded at the right moment. No overloaded context, no wasted tokens.

If AI agents interest you, I explained the phenomenon a few weeks ago. CLI tools are exactly what makes these agents useful in everyday work.

Bash: the real hero of the story

Anthropic's vision from the start was that models should know how to code. Coding means controlling a computer. And controlling a computer means using the terminal.

Bash is the universal tool. It can read files, write code, call APIs and filter results with head or grep… The agent can even limit the tokens it retrieves by piping commands. Try doing that with an MCP tool returning 10,000 tokens at once without filtering. Good luck.

Let's be honest: MCP was a good idea at the wrong time. We didn't yet know how to do this, so we created an abstraction. Now that bash and skills can do the job, that abstraction becomes a burden.

If you're still using MCP in 2026, it might be time for a clear-out. Your AI agent will thank you, and so will your token budget. Pretty good, right?

Sources: 10 CLI Tools That Make Claude Code UNSTOPPABLE — Chase (ChaseAI), STOP USING MCP: Here's what to use NOW (French video) — Melvyn (BeginReact)

Translated from the original French article. Publication dates, examples and figures refer to that original version.