Skip to content

Hosted MCP Onboarding Model

How-to Steps for one specific job. Assumes you already know the basics.

A product model for making this framework easier to adopt without softening it: a hosted MCP that handles onboarding, docs, prompts and visual help, plus a local MCP (or a thin local bridge) that does the actual writing in the user’s project.

Read this page when someone asks how the framework gets simpler without losing its teeth.

Related reference:

flowchart LR
A["User"] --> B["AI client"]
B --> C["Hosted SDD onboarding MCP"]
B --> D["Local sdd-mcp or bridge"]
C --> E["Docs, prompts, structure, examples"]
D --> F["Real project files"]
F --> G["idea/"]
F --> H["specs/"]
F --> I["bitacora/"]

Neither half works alone. A fully local MCP can write real files but still asks the user to install something, which is where most people quit. A fully hosted MCP connects in seconds and then cannot safely touch anything on the user’s disk.

So: hosting handles teaching, local handles execution. Onboarding gets easier without giving up real writes, and the rules stay the same across every AI client.

Purpose:

  • explain the framework
  • expose beginner prompts
  • expose visual folder maps
  • explain command outcomes
  • guide the user before any real project write happens

Recommended capabilities:

  • prompts like easy_start_project, easy_create_spec, easy_show_structure
  • static resources such as policy, quickstart, easy MCP guide, prompt packs
  • examples for new and existing projects
  • visual “what happens next” guidance

Purpose:

  • create folders and files
  • update specs/INDEX.md
  • write bitacora files
  • validate project state
  • check SDD gate before implementation

Recommended capabilities:

  • current sdd-mcp tools
  • optional wrapper CLI or desktop launcher for one-click local connection

For a non-technical user, the experience should feel like this:

flowchart TD
A["Connect one hosted MCP URL"] --> B["Ask in plain language"]
B --> C["AI explains what it will do"]
C --> D["Local MCP writes files only when needed"]
D --> E["User reviews and approves next step"]

The user should not need to understand:

  • transports
  • schemas
  • package builds
  • workspace rules in detail

The user should understand only:

  • what action is happening
  • what files will be touched
  • what result will appear
  • what the next step is

Short term:

  • keep sdd-mcp local for operations
  • publish docs that define the hosted layer contract
  • use the current HTTP transport as the conceptual base
  • optionally use GitMCP as the free external repo-context layer for public understanding

Mid term:

  • host a read-oriented onboarding MCP endpoint
  • expose prompts, easy guides, folder maps, and examples
  • keep writes local

Long term:

  • offer a one-click launcher or thin local bridge that the hosted layer can orchestrate through the client

GitMCP is useful as a free external layer for public repositories.

What it can do well:

  • let AI clients read and understand this repository remotely
  • expose public docs and structure with almost no setup
  • help with discovery, demos, and diffusion

What it should not be treated as:

  • a replacement for sdd-mcp
  • a replacement for your custom prompts and product behavior
  • a safe writer into the user’s local project

So the recommended framing is:

  • GitMCP or similar = external repo context
  • hosted onboarding MCP = your owned remote guidance layer
  • local sdd-mcp = operational local execution layer

The hosted onboarding MCP should provide at least:

  • sdd://docs/easy-mcp
  • sdd://docs/quickstart
  • sdd://policy/current
  • an easy command catalog
  • prompts for project start, spec creation, structure explanation, validation, next step, and session close
You can use SDD in two simple parts.
One part teaches and guides you.
The other part writes the real files in your project.
This keeps onboarding simple and execution safe.