CLI vs MCP choice mostly changes the HOW as a side effect. It doesn't answer the bigger question and probably harder one: who delegated the rigtht to cause that effect, for how long, and with what scope? Just like with people, you need a policy decision that's independent. It should be revocable and auditable.
One way that I look at it is with these long-running agents should look less like a script and more like an employee. You wouldn't give them the master key hoping they behave well. You'd give specific access and in stages probably. That's what I think we're missing with our agents is giving them appropriate authority, delegated by an owner with a audit trail
The debate around "MCP vs. CLI" is somewhat pointless to me personally. Use whatever gets the job done. MCP is much more than just tool calling - it also happens to provide a set of consistent rails for an agent to follow. Besides, we as developers often forget that the things we build are also consumed by non-technical folks - I have no desire to teach my parents to install random CLIs to get things done instead of plugging a URI to a hosted MCP server with a well-defined impact radius. The entire security posture of "Install this CLI with access to everything on your box" terrifies me.
The context window argument is also an agent harness challenge more than anything else - modern MCP clients do smart tool search that obviates the entire "I am sending the full list of tools back and forth" mode of operation. At this point it's just a trope that is repeated from blog post to blog post. This blog post too alludes to this and talks about the need for infrastructure to make it work, but it just isn't the case. It's a pattern that's being adopted broadly as we speak.
This has always surprised me as this always comes up in MCP discussions. To me, it just seem like a matter of updating the protocol to not have that context hungry behaviour. Doesn't seem like an insurmountable problem technically.
Glad you say it has already been addressed. Was the protocol itself updated to reflect that? Or are you just referring to off-spec implementations?
How, "Dynamic Tool Discovery"? Has this been codified anywhere? I've only see somewhat hacky implementations of this idea
https://github.com/modelcontextprotocol/modelcontextprotocol...
Or are you talking about the pressure being on the client/harnesses as in,
https://platform.claude.com/docs/en/agents-and-tools/tool-us...
If you don't change your approach but just use CLI "intead of" MCP, you'll end up with a new spin on the same problems. The guardrails MCP provides (identity, entitlement, multi-principal trust boundaries) still need to exist somewhere.
> The entire security posture of "Install this CLI with access to everything on your box" terrifies me This is fair for hosted MCPs, However I'm not claiming the CLI is universally more secure. users needs to know what they're doing.
Honestly though, after 20 years of this, the whole thread is debating the wrong layer. A well-designed API works through CLI, MCP, whatever. A bad one won't be saved by typed schemas.
> At this point it's just a trope that is repeated from blog post to blog post
Well, "Use whatever gets the job done" and "it's just a trope" can't both be true. If the CLI gets the job done for some use cases, it's not a trope. It's an option. And I'd argue what's happening is the opposite of a trope. Nobody's hyping CLIs because they're exciting. There's no protocol foundation, no spec committee, no ecosystem to sell into. CLIs are 40-year-old boring technology. When multiple teams independently reach for the boring tool, that's a signal, not a meme.
> This blog post too alludes to this and talks about the need for infrastructure to make it work
When tool search is baked into Claude Code, that's Anthropic building and maintaining the infrastructure for you. The search index, ranking, retrieval pipeline, caching. It didn't disappear. It moved.
And it only works in clients that support it. Try using tool search from a custom Python agent, a bash script, or a CI/CD pipeline. You're back to loading everything.
A CLI doesn't need the client to do anything special. `--help` works everywhere. That's the difference between infrastructure that's been abstracted away for some users and infrastructure that's genuinely not needed.
UNIX solved this with files and pipes for data, and processes for compute.
AI agents are solving this this with sub-agents for data, and "code execution" for compute.
The UNIX approach is both technically correct and elegant, and what I strongly favor too.
The agent + MCP approach is getting there. But not every harness has sub-agents, or their invocation is non-deterministic, which is where "MCP context bloat" happens.
Source: building an small business agent at https://housecat.com/.
We do have APIs wrapped in MCP. But we only give the agent BASH, an CLI wrapper for the MCPs, and the ability to write code, and works great.
"It's a UNIX system! I know this!"
Also interesting that while the big vendors are following this trend and are now trying to take a lead in it, they still suggest things like "but use a JSON schema" (the linked article does a bit of the same - acknowledging that incremental learning via `--help` is useful AND can be token-conserving (exception being that if they already "know" the correct pattern, they wouldn't need to use tokens to learn it, so there is a potential trade-off), they are also suggesting that LLMs would prefer to receive argument knowledge in json rather than in plain language, even though the entire point of an LLM is for understand and create plain language. Seemed dubious to me, and a part of me wondered if that advice may be nonsense motivated by desire to sell more token use. I'm only partially kidding and I'm still dubious of the efficacy.
* Here's a TL;DR for anyone who wants to skip the rest of this long message: I ran an LLM CLI eval in the form of a constructed CTF. Results and methodology are in the two links in the section linked: https://github.com/scottvr/jelp?tab=readme-ov-file#what-else
Anyhow... I had been experimenting with the idea of having --help output json when used by a machine, and came up with a simple module that exposes `--help` content as json, simply by adding a `--jelp` argument to any tool that already uses argparse.
In the process, I started testing, to see if all this extra machine-readable content actually improved performance, what it did to token use, etc. While I was building out test, trying to settle on legitimate and fair ways to come to valid conclusions, I learned of the OpenCLI schema draft, so I altered my `jelp` output to fit that schema, and set about documenting the things I found lacking from the schema draft, meanwhile settling to include these arg-related items as metadata in the output.
I'll get to the point. I just finished cleaning the output up enough to put it in a public repo, because my intent is to share my findings with the OpemCLI folks, in hopes that they'll consider the gaps in their schema compared to what's commonly in use, but at the same time, what came as a secondary thought in service of this little tool I called "jelp", is a benchmarking harness (and the first publishable results from it), the to me, are quite interesting and I would be happy if others found it to be and added to the existing test results with additional runs, models, or ideas for the harness, or criticism about the validity of the method, etc.
The evaluation harness uses constructed CLI fixtures arranged as little CLI CTF's, where the LLMs demonstrate their ability to use an unknown CLI be capturing a "flag" that they'll need to discover by using the usage help, and a trail of learned arguments.
My findings at first confirmed my intuitions, which was disappointing but unsurprising. When testing with GPT-4.1-mini, no manner of forcing them to receive info about the CLI via json was more effective than just letting them use the human-friendly plain English output of --help, and in all cases the JSON versions burned more tokens. I was able to elicit better performance by some measurements from 5.1-mini, but again the tradeoff was higher token burn.
I'll link straight to the part of the README that shows one table of results, and contains links to the LLM CLI CTF part of the repo, as well as the generated report after the phase-1 runs; all the code to reproduce or run your own variation is there (as well as the code for the jelp module, if there is any interest, but it's the CLI CTF eval that I expect is more interesting to most.)
https://github.com/scottvr/jelp?tab=readme-ov-file#what-else
The idea that people see this as one horn of a trilemma instead of just good practice is a bit strange. Who would complain that every import isn't a star-import? Bring in what you need at first, then load new things dynamically with good semantics for cascade / drill-down. Let's maybe abandon simple classics like namespacing and the unix philsophy for the kitchen-sink approach after the kitchen-sink thing is shown to work.
[1]: one might say 'of course you can just add details about the CLI to the prompt' ... which reinvents MCP in an ad hoc underspecified non-portable mode in your prompt.
The amortization point is interesting too. If you're running a support agent that calls the same 5 tools thousands of times a day, paying the schema cost once and caching it makes total sense. The post covers this in the "tightly scoped, high-frequency tools" section but your framing of it as a caching problem is cleaner.
On the footnote: guilty as charged, partially. The ~80 token prompt is a minimal bootstrap, not a full schema. It tells the agent how to discover, not what to call. But yeah, the moment you start expanding that prompt with specific flags and patterns, you're drifting toward a hand-rolled tool definition. The difference is where you stop. 80 tokens of "here's how to explore" is different from 10,000 tokens of "here's everything you might ever need." But the line between the two is blurrier than the post implies. Fair point.
Yes, MCP eats up context windows, but agents can also be smarter about how they load the MCP context in the first place, using similar strategy to skills.
The problem with tossing it out entirely is that it leaves a lot more questions for handling security.
When using skills, there's no implicit way to be able to apply policies in the sane way across many different servers.
MCP gives us a registry such that we can enforce MCP chain policies, i.e. no doing web search after viewing financials.
Doing the same with skills is not possible in a programatic and deterministic way.
There needs to be a middle ground instead of throwing out MCP entirely.
Are there any good docs youve liked to learn about it, or good open source projects you used to get familiar? I would like to learn more
@tool def do_great_thing(arg: string) -> string:
// todo
Notice that the tool uses structured inputs/outputs (the types - they can also be "dictionaries", or objects in most languages - giving the LLM powerful capabilities).
Now, imagine you want to write this in any language. What do you do?
Normally, you create some sort of API for that. Something like good old RPC. Which is essentially what MCP does: it defines a JSON-RPC API for tools, but it also adds some useful stuff, like access to static resources, elicitation (ask user for input outside of the LLM's chat) and since the MCP auth spec, an unified authorization system based on OAuth. This gives you a lot of advantages over a CLI, as well as some disadvantages. Both make sense to use. For example, for web usage, you just want the LLM to call Curl! No point making that a MCP server (except perhaps if you want to authorize access to URLs?). However, if you have an API that exposes a lot of stuff (e.g. JIRA) you definitely want a MCP for that. Not only does it get only the access you want to give the LLM instead of using your own credentials directly, now you can have a company wide policy for what can be done by agents when accessing your JIRA (or whatever) system.
A big disadvantage of MCP is that all the metadata to declare the RPC API take a lot of context, but recently agents are smart about that and load that partially and lazily as required, which should fix the problem.
In summary: whatever you do, you'll end up with something like MCP once you introduce "enterprise" users and not just yolo kids giving the LLM access to their browsers with their real credentials and unfiltered access to all their passwords.
You can think of an MCP server as a process exposing some tools. It runs on your machine communicating via stdin/stdout, or on a server over HTTP. It exposes a list of tools, each tool has a name and named+typed parameters, just like a list of functions in a program. When you "add" an MCP server to Claude Code or any other client, you simply tell this client app on your machine about this list of tools and it will include this list in its requests to the LLM alongside your prompt.
When the LLM receives your prompt and decides that one of the tools listed alongside would be helpful to answer you, it doesn't return a regular response to your client but a "tool call" message saying: "call <this tool> with <these parameters>". Your client does this, and sends back the tool call result to the LLM, which will take this into account to respond to your prompt.
That's pretty much all there is to it: LLMs can't connect to your email or your GitHub account or anything else; your local apps can. MCP is just a way for LLMs to ask clients to call tools and provide the response.
1. You: {message: "hey Claude, how many PRs are open on my GitHub repo foo/bar?", tools: [... github__pr_list(org:string, repo:string) -> [PullRequest], ...] } 2. Anthropic API: {tool_use: {id: 123, name: github__pr_list, input:{org: foo, repo: bar}}} 3. You: {tool_result: {id: 123, content: [list of PRs in JSON]} } 4. Anthropic API: {message: "I see 3 PRs in your repo foo/bar"}
that's it.
If you want to go deeper the MCP website[1] is relatively accessible, although you definitely don't need to know all the details of the protocol to use MCP. If all you need is to use MCP servers and not blow up your context with a massive list of tools that are included with each prompt, I don't think you need to know much more than what I described above.
web search is also another tool and you can gate it with logic so LLMs don’t go rogue.
that’s kinda simplest explanation i guess
Any kind of social push like that is always understood to be something to ignore if you understand why you need to ignore it. Do you agree that a typical solo dev caught in the MCP hype should run the other way, even if it is beneficial to your unique situation?
The only value in MCP is that it's intended "for agents" and it has traction.
I have been keeping an eye on MCP context usage with Claude Code's /context command.
When I ran it a couple months ago, supabase used 13.2k tokens all the time, with the search_docs tool using 8k! So, I disabled that tool in my config.
I just ran /context now, and when not being used it uses only ~300 tokens.
I have a question. Does anyone know a good way to benchmark actual MCP context usage in Claude Code now? I just tried a few different things and none of them worked.
The people saying this and attacking it should first agree about the question.
Are you combining a few tools in the training set into a logical unit to make a cohesive tool-suite, say for reverse engineering or network-debugging? Low stakes for errors, not much on-going development? Great, you just need a thin layer of intelligence on top of stack-overflow and blog-posts, and CLI will probably do it.
Are you trying to weld together basically an AI front-end for an existing internal library or service? Is it something complex enough that you need to scale out and have modular access to? Is it already something you need to deploy/develop/test independently? Oops, there's nothing quite like that in the training set, and you probably want some guarantees. You need a schema, obviously. You can sort of jam that into prompts and prayers, hope for the best with skills, skip validation and risk annotations being ignored, trust that future opaque model-change will be backwards compatible with how skills are even selected/dispatched. Or.. you can use MCP.
Advocating really hard for one or the other in general is just kind of naive.
MCPs are clunky, difficult to work with and token inefficient and security orgs often have bad incentive design to mostly ignore what the business and devs need to actually do their job, leading to "endpoint management" systems that eat half the system resources and a lot of fig leaf security theatre to systematically disable whatever those systems are doing so people can do their job in an IT equivalent that feels like the TSA.
Thank god we moving away from giving security orgs these fragile tools to attach ball and chains to everyone.
Here's a longer piece on why the trust boundary has to live at the runtime level, not the interface level, and what that means for MCP's actual job: https://forestmars.substack.com/p/twilight-of-the-mcp-idols
Do you have some more info on it?
looking up "registry" in the mcp spec will just describe a centrally hosted, npm-like package registry[^1]
[^1]: The MCP Registry is the official centralized metadata repository for publicly accessible MCP servers, backed by major trusted contributors to the MCP ecosystem such as Anthropic, GitHub, PulseMCP, and Microsoft.
With CLI, it's your machine, your keys. With direct API calls, keys live wherever the agent runs. Both work until a contractor leaves and their laptop still has active keys for your repos, your internal docs, and your CRM.
Remote MCP over streamable HTTP gives you a centralized auth layer. One SSO integration, one revocation point, one audit trail.
I wrote about this angle here: https://dev.to/dennistraub/missing-from-the-mcp-debate-who-h...
is this Human 2.0? I only have 1.0a beta in the office.
I get the joke but it really does highlight how flimsy the argument is for humans. IME humans frequently make simple errors everywhere they don’t learn from and get things right the first time very rarely. Damn. Sounds like LLMs. And those are only getting better. Humans aren’t.
Terminator 2 Clip: https://youtu.be/XTzTkRU6mRY?t=72&si=dmfLNDqpDZosSP4M
However, MCPs have some really nice properties that CLIs generally don’t, or that are harder to solve for. Most notably, making API secrets available to the CLI, but not to the agent, is quite tricky. Even in this example, the options are env variables (which are a prompt injection away from dumping), or a credentials file (better, but still very much accessible to the agent if it were asked).
MCPs give you a “standard” way of loading and configuring a set of tools/capabilities into a running MCP server (locally or remotely), outside of the agent’s process tree. This allows you to embed your secrets in the MCP server, via any method you choose, in a way that is difficult or impossible for the agent to dump even if it goes rogue.
My efforts to replicate that secure setup for a CLI have either made things more complicated (using a different user for running CLIs so that you can rely upon Linux file permissions to hide secrets), or start to rhyme with MCP (a memory-resident socket server started before the CLI that the CLI can talk to, much like docker.sock or ssh-agent)
Much easier:
{ action: 'help' }
{ action: 'projects.help' }
{ action: 'projects.get', payload: { id: xxxx-xx-x } }
There are other interesting capabilities though, like built in permissions based on HTTP verb, that might be useful to someone.
The fix I (well Codex actually) landed on was toolset tiers (minimal/authoring/experimental) controlled by env var, plus phase-gating, now tools are registered but ~80% are "not connected" until you call _connect. The effective listed surface stays pretty small.
Lazy loading basically, not a new concept for people here.
The major harnesses like Claude Code + Codex have had tool search for months now.
But tool search is solving the symptom, not the cause. You still pay the per-tool token cost for every tool the search returns. And you've added a search step (with its own latency and token cost) before every tool call.
With a CLI, the agent runs `--help` and gets 50-200 tokens of exactly what it needs. No search index, no ranking, no middleware. The binary is the registry.
Tool search makes MCP workable. CLIs make the search unnecessary.
Wait, better check help. is it -h? [error]
Nope? Lemme try —-help. [error]
Nope.
How about just “help” [error]
Let me search the web [tons of context and tool calls]
Very interesting topic, but this LLM structure is instant anthema I just have to stop reading once I smell it.
The fix that worked for us was giving agents a CLI instead. ~80 tokens in the system prompt, progressive discovery through --help, and permission enforcement baked into the binary rather than prompts.
The post covers the benchmarks (Scalekit's 75-run comparison showed 4-32x token overhead for MCP vs CLI), the architecture, and an honest section on where CLIs fall short (streaming, delegated auth, distribution).
Compare this to an MCP, where my understanding is that the entire API usage is injected into the context.
Why not run the discovery (whether MCP or CLI) in a subagent that returns only the relevant tools. I mean, discovery can be done on a local model, right?
This might be a complete non issue in 6 months.
The pattern with every resource expansion is the same: usage scales to fill it. Bigger windows mean more integrations connected, not leaner ones. Progressive disclosure is cheaper at any window size.