llm-kms-mcp 0.1.0

MCP Resource Server — llm-kms handbook distributor

llm-kms MCP Handbook Server

An MCP Resource Server that ships the llm-kms schema, skill, and templates as a self-contained package. Any agent can read these resources to bootstrap its own llm-kms wiki in any repo — no external repo required.

What it exposes

Resource URI	Content
llmkms://guide/bootstrap	Step-by-step bootstrap guide (start here)
llmkms://spec/claude-md	Full authoritative schema (CLAUDE.md)
llmkms://skill/skill-md	Thin router (SKILL.md) — points to the per-op skills below
llmkms://skill/ingest	Ingest procedure
llmkms://skill/query	Query procedure
llmkms://skill/lint	Lint procedure
llmkms://skill/prune	Prune procedure
llmkms://template/atomic-note	Blank atomic-note template
llmkms://template/index	Blank index.md starter
llmkms://template/log	Blank log.md starter
llmkms://reference/gitignore	Recommended .gitignore
llmkms://reference/gitleaks	Recommended .gitleaks.toml
llmkms://reference/lint-script	Standalone Python lint script
llmkms://manifest	JSON index with sha256 per resource

All resources are bundled inside the package — no environment variables needed.

Tools

Tool	Parameters	Does
scaffold_llm_kms	target_dir	Scaffolds an llm-kms wiki folder structure (inbox/, raw/, archive/, system/templates/, index.md, log.md, merged .gitignore) in target_dir. Safe to re-run — skips files that already exist.
scaffold_mcp_server	target_dir, package_name, port=8765, concurrent_writes=False	Generates a standalone wiki MCP server project in target_dir: core.py, server_public.py (HTTP, query-only), server_local.py (stdio, full tool set), a generated _config.py with a random bearer token, pyproject.toml, and a copy of the per-op skill files under .claude/skills/<package_name>/. concurrent_writes=True switches generated note IDs from 17 to 25 digits (collision-safe for multiple agents writing at once).

Requirements

• Python >= 3.10

Install

From PyPI

pip install llm-kms-mcp

From source

git clone https://github.com/operator47/llm-kms-mcp.git
cd llm-kms-mcp
pip install .

Run (smoke test with MCP Inspector)

Node.js must be installed.

npx @modelcontextprotocol/inspector python -m llm_kms_mcp

In the Inspector: confirm all 14 resources appear under llmkms:// and both tools (scaffold_llm_kms, scaffold_mcp_server) are listed; read llmkms://guide/bootstrap to verify content.

Register with Claude Code

Project-scoped (.mcp.json in your repo):

{
  "mcpServers": {
    "llm-kms-handbook": {
      "command": "python",
      "args": ["-m", "llm_kms_mcp"]
    }
  }
}

With uv (no separate install step):

{
  "mcpServers": {
    "llm-kms-handbook": {
      "command": "uvx",
      "args": ["llm-kms-mcp"]
    }
  }
}

User-scoped via CLI:

claude mcp add llm-kms-handbook -- python -m llm_kms_mcp

Acceptance checks

After installing:

1. All 14 resources are listed in the Inspector.
2. llmkms://guide/bootstrap returns the bootstrap guide.
3. llmkms://manifest lists all resources with sha256 hashes and no errors.
4. A new empty repo bootstrapped from the guide produces a clean lint result.
5. scaffold_llm_kms and scaffold_mcp_server are both callable and produce the files described above.