First, MCP tools are sent on every request. If you look at the notion MCP the search tool description is basically a mini tutorial. This is going right into the context window. Given that in most cases MCP tool loading is all or nothing (unless you pre-select the tools by some other means) MCP in general will bloat your context significantly. I think I counted about 20 tools in GitHub Copilot VSCode extension recently. That's a lot!

Second, MCP tools are not compossible. When I call the notion search tool I get a dump of whatever they decide to return which might be a lot. The model has no means to decide how much data to process. You normally get a JSON data dump with many token-unfriendly data-points like identifiers, urls, etc. The CLI-based approach on the other hand is scriptable. Coding assistant will typically pipe the tool in jq or tail to process the data chunk by chunk because this is how they are trained these days.

If you want to use MCP in your agent, you need to bring in the MCP model and all of its baggage which is a lot. You need to handle oauth, handle tool loading and selection, reloading, etc.

The simpler solution is to have a single MCP server handling all of the things at system level and then have a tiny CLI that can call into the tools.

In the case of mcpshim (which I posted in another comment) the CLI communicates with the sever via a very simple unix socket using simple json. In fact, it is so simple that you can create a bash client in 5 lines of code.

This method is practically universal because most AI agents these days know how to use SKILLs. So the goal is to have more CLI tools. But instead of writing CLI for every service you can simply pivot on top of their existing MCP.

This solves the context problem in a very elegant way in my opinion.

> Before your agent can do anything useful, it needs to know what tools are available. MCP’s answer is to dump the entire tool catalog into the conversation as JSON Schema. Every tool, every parameter, every option.

Because this simply isn't true anymore for the best clients, like Claude Code.

Similar to how Skills were designed[1] to be searchable without dumping everything into context, MCP tools can (and does in Claude Code) work the same way.

See https://www.anthropic.com/engineering/advanced-tool-use and https://x.com/trq212/status/2011523109871108570 and https://platform.claude.com/docs/en/agents-and-tools/tool-us...

[1] https://agentskills.io/specification#progressive-disclosure

The CLI approach definitely has practical benefits for token reduction. Not stuffing the entire schema into the runtime context is a clear win. But my main interest lies less in "token cost" and more in "how we structure the semantic space."

MCP is fundamentally a tool-level protocol. Existing paradigms like Skills already mitigate context bloat and selection overhead pretty well via tool discovery and progressive disclosure. So framing this purely as "MCP vs CLI" feels more like shifting the execution surface rather than a fundamental architectural shift.

The direction I'm exploring is a bit different. Instead of treating tools as the primary unit, what if we normalize the semantic primitives above them (e.g., "search," "read," "create")? Services would then just provide a projection of those semantics. This lets you compress the semantic space itself, expose it lazily, and only pull in the concrete tool/CLI/MCP adapters right at execution time.

You can arguably approximate this with Skills, but the current mental model is still heavily anchored to "tool descriptions"—it doesn't treat normalized semantics as first-class citizens. So while the CLI approach is an interesting optimization, I'm still on the fence about whether it's a real structural paradigm shift beyond just saving tokens.

Ultimately, shouldn't the core question be less about "how do we expose fewer tools," and more about "how do we layer and compress the semantic space the agent has to navigate?"

I do understand anthropic's Tool Search helps with mcp bloat, but it's limited only to claude.

CMCP currently supports codex and claude but PRs are welcome to add more clients.

[1]https://blog.cloudflare.com/code-mode-mcp/ [2]https://github.com/assimelha/cmcp

Awesome TUIs: https://github.com/rothgar/awesome-tuis

Awesome CLIs: https://github.com/agarrharr/awesome-cli-apps

Terminal Trove: https://terminaltrove.com/

I guess this is another one shows that the CLI and Unix is coming back in 2026.

I've also launched https://mcpshim.dev (https://github.com/mcpshim/mcpshim).

The unix way is the best way.

This way you build all your MCPs into the system prompt, save the prompt to the AI provider, then use it without overpaying API costs.

The current "tools-on-demand" workarounds should be great for infrequent tools but the future will probably bring agents with dozens of tools that need them in context to flexibly many of them in the same context window. So we just need to make the context windows longer and make this capability cheaper to use.

One thing I have read recently is that when you make a tool call it forces the model to go back to the agent. The effect of this is that the agent then has to make another request with all of the prompt (include past messages), these will be "cached" tokens, but they're still expensive. So if you can amortize the tool calls by having the model either do many at once or chaining them with something like bash you'll be better off.

I suspect this might be why cursor likes writing bash scripts so much, simple shell commands are going to be very token heavy because of the frequency of interrupts.

MCP's token cost is the price of availability. The fix isn't to replace the protocol, it's to only activate the tools that matter for the current context. Claude's Skills already work this way -> lightweight descriptions loaded upfront, full definitions fetched on demand. That's essentially the same lazy-loading pattern CLIHub describes, just built into the model's native workflow.

But yeah, a concrete example is playwright-mcp vs playwright-cli: https://testcollab.com/blog/playwright-cli

For other integrations, I first try to find an official or unofficial CLI tool (a wrapper around the API), and only then do I consider using MCP

https://jannikreinhard.com/2026/02/22/why-cli-tools-are-beat...

Even the smallest models are RL trained to use shell commands perfectly. Gemini 3 flash performs better with a cli with 20 commands vs 20+ tools in my testing.

cli also works well in terms of maintaining KV cache (changing tools mid say to improve model performance suffers from kv cache vs cli —help command only showing manual for specific command in append only fashion)

Writing your tools as unix like cli also has a nice benefit of model being able to pipe multiple commands together. In the case of browser, i wrote mini-browser which frontier models use much better than explicit tools to control browser because they can compose a giant command sequence to one shot task.

https://github.com/runablehq/mini-browser

https://blog.cloudflare.com/code-mode-mcp/

https://news.ycombinator.com/item?id=47129241

[0] https://github.com/steipete/mcporter

ALSO... the permission boundary is clearer. You can whitelist commands, flags, working dir... it becomes manageable.

HOWEVER... packaging still matters. A “small” CLI that pulls in a giant runtime kills the benefit.

I want the discipline of small protocol plus big cache. Cheap models can summarize what they did and avoid full context in every step...

Maybe MCP can help segregate auto-approve vs ask more cleanly, but I don't actually see that being done.

The real killer is the input tokens on each step. If you have 100k tokens in the conversation, and the LLM calls an MCP tool, the output and the existing conversation is sent back. So now you've input 200k tokens to the LLM.

Now imagine 10 tool calls per user message - or 50. You're sending 1-5M input tokens, not because the MCP definitions or tool responses are large, but because at each step, you have to send the whole conversation again.

"what about caching" - Only 90% savings, also cache misses are surprisingly common (we see as low as 40% cache hit rate)

"MCP definitions are still large" - not compared to any normal conversation. Also these get cached

We've seen the biggest savings by batching/parallelizing tool calls. I suspect the future of LLM tool usage will have a different architecture, but CLI doesn't solve the problems either.

[0] https://ziva.sh, it's an agent specialized for Godot[1]

[1] https://godotengine.org

But MCP today isn’t ideal. I think we need to have some catalogs where the agents can fetch more information about MCP services instead of filling the context with not relevant noise.

I can see a future where software is built with a CLI interface underneath the (optional) GUI, letting an LLM hook directly into the underlying "business" logic to drive the application. Since LLM's are basically text machines, we just need somebody to invent a text-driven interface for them to use...oh wait!

Imagine booking a flight - the LLM connects to whatever booking software, pulls a list of commands, issues commands to the software, and then displays the output to the user in some fashion. It's basically just one big language translation task, something an LLM is best at, but you still have the guardrails of the CLI tool itself instead of having the LLM generate arbitrary code.

Another benefit is that the CLI output is introspectable. You can trace everything the LLM is doing if you want, as well as validate its commands if necessary (I want to check before it uses my credit card). You don't get this if it's generating a python script to hit some API.

Even before LLM's developers have been writing GUI applications as basically a CLI + GUI for testability, separation of concerns etc. Hopefully that will become more common.

Also this article was obviously AI generated. I'm not going to share my feelings about that.

I know I saw something about the Next.js devs experimenting with just dumping an entire index of doc files into AGENTS.md and it being used significantly more by Claude than any skills/tool call stuff.

The biggest difference is state, but that's also kind of easy from CLI, the tool just have to store it on disk, not in process memory.

Edit: took out because I think that was something different.