MCP is very much not dead. centralized remote MCP servers are incredibly useful. also bespoke CLIs still require guidance for models to use effectively, so it's clear that token efficiency is still an issue regardless.
MCP is very much not dead. centralized remote MCP servers are incredibly useful. also bespoke CLIs still require guidance for models to use effectively, so it's clear that token efficiency is still an issue regardless.
Tbh I find self-documenting CLIs (e.g. with a `--help` flag, and printing correct usage examples when LLMs make things up) plus a skill that's auto invoked to be pretty reliable. CLIs can do OAuth dances too just fine.
MCP's remaining moats I think are:
- No-install product integrations (just paste in mcp config into app)
- Non-developer end users / no shell needed (no terminal)
- Multi-tenant auth (many users, dynamic OAuth)
- Security sandboxing (restrict what agents can do), credential sandboxing (agents never see secrets)
- Compliance/audit (structured logs, schema enforcement)?
If you're a developer building for developers though, CLI seems to be a clear winner right
Imagine if, in addition to local MCP "servers", the MCP people had nurtured a structured CLI-based --help-equivalent consumable by LLMs and shell completion engines alike. Doing so, you unify "CLI" (trivial deployment; human accessibility) and MCP-style (structured and discoverable tool calling) in a single DWIM artifact.
But since when has this industry done the right thing informed by wisdom and hindsight?
that's a pretty interesting idea. It would be nice if there was such a standard. the approach I'm taking right now: a CLI that accepts structured JSON as input, with an 'mcp' subcommand that starts a stdio server. I bundle a 'help' command with a 'describe' action for self-service guidance scoped to a particular feature/tool.
There are actually a lot of great things you can to to make CLIs more helpful to agents. I use a structured help called '--capabilities' but there is a ton of JIT context you can do from the CLI as well https://keyboardsdown.com/posts/01-agent-first-clis/
That structured CLI already exists: PowerShell cmdlets.
Not in any meaningful, general way on Unix systems it doesn't. Nobody uses psh outside Windows
But nobody is using your hypothetical "structured CLI-based --help-equivalent consumable by LLMs and shell completion engines alike" either. In terms of mindshare, you're starting from scratch either way.
I just remembered docopt, which maybe fits the bill in a more Unixy way, but it and its ports are mostly abandoned, for various reasons.
Maybe they should consider using it.
Okay, but they won't. We both know that. Now what?
I see remote MCP servers as a great interface to consume api responses. The idea that you essentially make your apis easily available to agents to bring in relevant context is a powerful one.
When folks say MCP is dead, I don't get it. What other alternatives exist in place of MCP? Arbitrary code via curl/sdks to call a remote endpoint?
> What other alternatives exist in place of MCP? Arbitrary code via curl/sdks to call a remote endpoint?
cli?
for example aws cli. It's a full interface to aws API. Why would you need mcp for that?
and if you have any doubts, agents use it with a great effect even without any relevant skill. "aws help" is fully discoverable.
yes, but clis thus need self-service commands to provide guidance, and their responses need to be optimized for consumption by agents. in a sense, this is the same sort of context tax that MCP servers incur. so in my view cli and MCP are complementary tools; one is not strictly superior over the other.
> yes, but clis thus need self-service commands to provide guidance, and their responses need to be optimized for consumption by agents.
MCP vs Agent Skills:
MCPs once configured cost you tokens even when they are not used. Unlike MCPs, skills use progressive disclosure. The AI agent does not load up the entire context, if the skill is not being used.
MCPs will die off mostly for this reason alone.
I think cli’s are more token efficient- the help menu is loaded only when needed, and the output is trivially pipe able to grep or jq to filter out what the model actually wants
all you need is a simple skills.md and maybe a couple examples and codex picks up my custom toolkit and uses it.
whats your custom toolkit
I have dozens of clis that are custom built for codex to use.