MCP/skillgen

• mcp-skillgen

  • Abstract idea

    • Let mytech-mcp be an existing Open Source MCP/Server.
    • Let mcp-skillgen be a CLI/Tool, or more appropriately, a CLI metatool, which facilitates the transformation of mytech-mcp into a dedicated mytech-mcp-cli CLI tool, and which also facilitate the use of that generated CLI to extract and instantiate subsets of functionality into dedicated Claude Code Skills.

  • Concrete idea

    • There's a Claude Code skill called mcp-skillgen which wraps a CLI tool of the same name.
    • Let's say mytech-mcp contains 70 tools. You only need 3 - tool1, tool2, tool3. You call either instruct Claude to use the skill, or directly call
      • mcp-skillgen path/to/mytech-mcp/source --skillgen tool1 --skillgen tool2, --skillgen tool3
    • Result
      • A mytech-mcp-cli is created
        • which reimplements the functionality of tool1, tool2, tool3
        • which knows that it has not compiled tool4 through tool70, and can answer questions about those tools but not execute them
        • which knows how to use mcp-skillgen to compile any of the tools, if they needed
      • A Claude Code Project Skill is created called mytech which wraps mytech-mcp-cli; it knows all the things mytech-mcp-cli knows and can do
      • A Claude Code Project Skill is created for mytech-tool1, mytech-tool2, mytech-tool3, which tells it to use the generated mytech-mcp-cli to accomplish the skill. mytech-mcp-cli is located in .claude/skills/mytech/resources/mytech-mcp-cli
    • Now
      • Claude code knows how to use the skills from the MCP without weighing down the context.

  • Impetus

    • MCP is extremely wasteful. AI/Context/Rot is a very real problem, and MCPs exacerbate it.
    • If an MCP server has 70 tools, it's quite likely that any given conversation will not need more than three of those tools. So 57 tool definitions clog the entire context. It's quite common for 2-3 MCP servers to consume > 40% of the context. Add Anthropic, GitHub and one more, and it will be at 40% of the context limit for Anthropic/Model/Claude/Sonnet/4.5.
    • MCP also loads tool definitions into context long before they are used. If there are ten turns to a conversation and the tool is needed on the 11th, then loading those tools before the 11th is wasteful.

  • Elaboration and Questions

    • Why generate a CLI for replacing the MCP/Server?

      • Any good CLI will have documentation. By creating a dedicated CLI that reimplements the functionality of the MCP server, an AI agent will be able to use its documentation and help pages at the command line to discover how to use it.

    • Why use a CLI to generate a CLI?

      • In Agentic AI Coding agents, CLIs are more context efficient than MCP, for the reasons listed in the Impetus. As a result, the same holds true for
      • coding agents know how to use CLI tools; we don't need to reinvent the world there, and also they don't need to be spoon-fed instructions about how to use the tools before they are needed.
      • The reason why we want to use generated CLIs are the reason why we want to use a CLI to generate them.

    • How will the CLI be generated?

      • Given the source code for the MCP, mcp-skillgen will wrap commands to a CLI-based AI/Coding/Tool to
        • file information about the tool in a file for later
          • one-time setup
            • analyze the MCP architecture of the tool with respect to security, authentication and other NFRs
          • some setup for all tools, whether implemented or not
            • create a basic requirements spec for the tool's basic functional requirements
          • some setup for all implemented tools
            • create a detailed requirements spec for the tool's detailed functional requirements (for implemented tools)
            • analyze the implementation and enhance the spec with key lessons learned
        • implementation
          • implement the CLI targets and their corresponding skills