One thing that MCP solves well, that neither CLI apps (like the `gh` CLI for example) nor letting your LLM call arbitrary APIs via CURL does, is setting granular permissions per tool.
Most agent frontends I've used like Claude Code only give you one level deep of CLI commands to authorize, which works fine for allowing commands like `docker build:*`. But for complex CLIs like GitHub, Azure, etc. it just doesn't scale well. It is absurd to grant Claude Code permission to `az vm:*` when that includes everything from `az vm show` to `az vm delete`. Likewise, the argument that says that you should just let your LLM call APIs directly via curl or whatever, does not hold up well when Claude Code just wants raw access to all of `curl:*`.
Meanwhile, MCP tools are (currently, at least in CC) managed at the individual tool level, which is very convenient for managing granular permissions.
Perhaps there could be some "CTCP" (CLI tool context protocol; the CCP acronym does not work well) where CLI apps could expose their available tools to the LLM, and it could then be dynamically loaded and managed at a granular level. But until then, I'm going to keep using MCP.
Most agent frontends I've used like Claude Code only give you one level deep of CLI commands to authorize, which works fine for allowing commands like `docker build:*`. But for complex CLIs like GitHub, Azure, etc. it just doesn't scale well. It is absurd to grant Claude Code permission to `az vm:*` when that includes everything from `az vm show` to `az vm delete`. Likewise, the argument that says that you should just let your LLM call APIs directly via curl or whatever, does not hold up well when Claude Code just wants raw access to all of `curl:*`.
Meanwhile, MCP tools are (currently, at least in CC) managed at the individual tool level, which is very convenient for managing granular permissions.
Perhaps there could be some "CTCP" (CLI tool context protocol; the CCP acronym does not work well) where CLI apps could expose their available tools to the LLM, and it could then be dynamically loaded and managed at a granular level. But until then, I'm going to keep using MCP.