HVE Guidance MCP connector for Copilot Studio

April 20, 2026

AI engineering standards are only useful if developers can find them, apply them, and validate their work against them. The microsoft/hve-core repository holds curated guidance for AI engineering—instructions, prompts, skills, and agent configurations—but browsing a GitHub repository is not a natural part of a Copilot Studio conversation.

This connector bridges that gap. Twelve MCP tools let your agent discover assets, validate content quality, track changes, and generate adoption plans, all without leaving the conversation.

Full source: GitHub repository

What the connector covers

The 12 tools group into four areas:

Discovery

Validation

Change intelligence

Workflow and adoption

How the tools work together

The tools compose naturally in a conversation. An agent can:

1. Call recommend_assets_for_task to find relevant assets for “add governance to our agent pipeline”
2. Call get_asset on the top result to see intended use, constraints, and related assets
3. Call validate_instruction against a draft instruction to catch issues before merging
4. Call generate_adoption_plan to produce a phased rollout for the team

The recommendation tool goes beyond path matching. It scores candidates using task intent labels, frontmatter metadata, and summary matching, so “add code review standards” doesn’t return generic onboarding assets.

Validation tools return structured findings with three fields per issue: what the problem is, why it matters, and how to fix it. That structure is more useful to an agent than a plain warning string.

Asset types

The connector maps asset types to their standard HVE repository paths:

Example prompt flows

Discover and review

User: What HVE assets should I use for a risky multi-file refactor?
Agent: [recommend_assets_for_task] → planning, research, review assets ranked by relevance
Agent: [get_asset] → opens top asset, shows constraints and related assets

Validate before merging

User: Validate this instruction and tell me what to fix.
Agent: [validate_instruction] → returns findings with message, why, and fix per issue

Team rollout planning

User: Create an 8-week adoption plan for the API Platform team to standardize agent prompts.
Agent: [generate_adoption_plan] → phased milestones, relevant tool suggestions, team-scoped goals

Track what changed

User: Summarize HVE instruction changes from the last 30 days.
Agent: [summarize_asset_changes] → commit list scoped to instruction paths
User: Did the code review instruction change materially since last month?
Agent: [compare_asset_versions] → diff summary with changed status and estimated line deltas

Runtime details

The connector reads from GitHub using the REST API with short-lived in-memory caching (5 minutes for most operations, 2 minutes for search) to reduce duplicate requests and lower latency. GitHub throttling errors return structured retry metadata rather than unhandled exceptions.

Tool errors return as MCP tool results with isError: true, so the agent can handle failures gracefully without breaking the conversation.

Authentication

Connection parameters:

Use a read-only personal access token scoped to repository access. Pointing the connector at a fork or internal mirror changes only the connection parameters—the tool logic stays the same.

Deploy

pac auth create --environment "https://yourorg.crm.dynamics.com"

pac connector create `
  --api-definition-file apiDefinition.swagger.json `
  --api-properties-file apiProperties.json `
  --script-file script.csx

If --script-file fails in your environment, upload script.csx manually in the connector Code tab after creation.

Evaluation pack

The repository includes docs/evaluation-scenarios.md with 14 manual scenarios covering discovery, retrieval, recommendation, validation, change intelligence, and adoption planning. Each scenario includes expected tool selection and scoring criteria across tool correctness, result usefulness, explanation quality, ranking, and actionability.

Use it to track quality regressions when the upstream HVE repository changes or when you swap the connected repository.

Observability

Add your Application Insights instrumentation key to script.csx to enable telemetry:

private const string APP_INSIGHTS_KEY = "your-instrumentation-key";

Logged events include McpRequestReceived, McpRequestCompleted, and per-tool exceptions with correlation IDs across all events.

Files

Resources

• HVE Guidance connector source code
• microsoft/hve-core repository
• Custom connectors overview
• Power Platform CLI overview
• Application Insights overview