Power Compendium MCP connector for Power Platform

April 14, 2026

RAG re-reads your documents on every query. It retrieves chunks, feeds them to an LLM, and hopes the answer lands somewhere in the context window. It works, but it’s wasteful—the LLM re-derives the same knowledge every time someone asks a similar question.

Karpathy’s LLM Wiki pattern flips this. Instead of retrieval-at-query-time, the LLM processes documents once, extracts entities and concepts into interlinked pages, and maintains that knowledge base over time. Queries search the compendium, not the raw documents.

Power Compendium implements this pattern as an Azure Container App with a Power Platform connector—11 operations accessible via MCP (Copilot Studio, M365 Copilot, VS Code) and REST (Power Automate, Power Apps).

Full source: GitHub repository

How it differs from RAG

11 operations

Book operations

Page management

Three ways to ingest knowledge

Single source

Feed in articles, meeting notes, API docs, papers—one document at a time. The LLM reads it, extracts key information, creates or updates relevant pages, and maintains cross-references.

POST /api/book/ingest?scope=org
{
    "title": "Rubberducking with LLMs",
    "content": "Rubber duck debugging is a method where...",
    "sourceUrl": "https://deepgram.com/learn/llms-the-rubber-duck-debugger",
    "category": "article"
}

Multi-file skill

Ingest an entire agent skill—instruction files, templates, and examples processed together. The LLM understands the relationships between files and extracts richer cross-references than individual ingestion.

POST /api/book/ingest-skill?scope=org
{
    "skillName": "mcp-connector",
    "files": [
        {"path": "SKILL.md", "content": "# MCP Connector for Copilot Studio..."},
        {"path": "template/script.csx", "content": "public class Script : ScriptBase..."},
        {"path": "template/apiDefinition.swagger.json", "content": "{\"swagger\":\"2.0\",...}"}
    ]
}

URL-based fetch

Point the compendium at a URL and it fetches the content automatically. GitHub repository tree URLs download all Markdown, code, JSON, and YAML files in the directory.

POST /api/book/ingest-from-url?scope=org
{
    "url": "https://github.com/troystaylor/SharingIsCaring/tree/main/.github/skills/mcp-connector",
    "type": "agent-skill"
}

The ingestion pipeline

Regardless of which option you use, the LLM processing pipeline is the same:

1. Read — The LLM reads the source content and the current book index
2. Plan — It determines what pages to create (entities, concepts, source summaries) and what existing pages to update
3. Write — New pages are created and existing pages are updated with new information
4. Cross-reference — Pages link to each other using [[page-id]] syntax
5. Index — All new/updated pages are indexed in Azure AI Search
6. Log — A chronological entry is appended to the book’s activity log

Scoping: personal and org books

Every operation accepts a scope parameter:

• org (default) — shared organizational book
• personal — your private book (keyed to your Entra ID)
• all — search across both (query and list only)

The promote_page tool copies a page from your personal book to the org book—useful when you’ve curated something worth sharing with the team.

Page categories

Lint: knowledge base health checks

The lint_book tool runs a health check across the entire compendium—detecting contradictions between pages, stale claims, orphan pages with no cross-references, missing cross-references, and knowledge gaps. Think of it as a linter for your organizational knowledge.

Architecture

M365 Copilot / VS Code / Claude Desktop
    ↓ (MCP, native)
                                        Azure Container App
Copilot Studio / Power Automate             ↓
    ↓ (MCP via connector)              ASP.NET Core API
    Power Compendium Connector ──→          ├── BookService
                                            ├── BookStorageService (Azure Blob)
                                            ├── BookSearchService (Azure AI Search)
                                            ├── BookLlmService (Azure OpenAI)
                                            └── EasyAuth (Entra ID OAuth v2)

Azure resources

Use cases

Organizational knowledge base: Ingest internal documentation, architecture decisions, runbooks, and tribal knowledge. Teams query the compendium through a Copilot Studio agent instead of searching SharePoint or asking the person who “just knows.”

Agent skill library: Use ingest_skill and ingest_from_url to build a searchable compendium of all your agent skills—instructions, templates, and patterns. New team members ask the agent “how do we build an MCP connector?” and get a synthesized answer with cross-references to related skills.

Continuous learning agent: A Power Automate flow triggers ingest_source whenever a new document lands in a SharePoint library. The compendium grows automatically, and the lint_book tool flags when new information contradicts existing pages.

Personal research assistant: Use the personal scope to build your own knowledge base. Ingest articles, papers, and notes. Query your compendium for synthesized answers with citations. Promote useful pages to the org book when they’re ready.

Security

• EasyAuth validates Entra ID tokens before requests reach application code
• Managed Identity for all Azure service access (Blob, Search, OpenAI)—no keys stored
• Input validation—pageId allowlist (kebab-case only), content size limits (100 KB), question limits (1 KB)
• Rate limiting—30 requests/minute per user on LLM-bound endpoints
• Log sanitization—user-provided content stripped of Markdown injection before logging

Deployment

Deploy the Azure infrastructure

.\infra\deploy.ps1 `
    -ResourceGroup rg-power-compendium `
    -OpenAiResourceGroup myResourceGroup `
    -OpenAiAccountName my-openai-resource `
    -OpenAiDeploymentName gpt-4o

The deploy script provisions all Azure resources (Storage, AI Search, ACR, Container Apps), builds and pushes the container image, assigns RBAC, and verifies the deployment.

Deploy the connector

pac connector create `
    --api-def connector/apiDefinition.swagger.json `
    --api-prop connector/apiProperties.json `
    --script connector/script.csx

Sign in with your Entra ID account when creating a connection. No additional configuration needed.

Resources

• Karpathy’s LLM Wiki pattern
• Azure AI Search documentation
• Azure OpenAI documentation
• Full source code