MCP Deep Dive: Five Servers, Five Reasoning Skills

OpenClaw Academy · Part 2, Issue 10 · Deep dive

Issue 09 established the architecture: skills teach reasoning, MCP servers provide capability. This issue implements it. Five servers, each with its own companion skill that demonstrates the correct pattern — instructions that teach when and how to use the server’s tools, not instructions that make API calls directly.

---

The difference between a skill that uses MCP and one that doesn’t

Here is the same task implemented two ways.

Wrong — skill with embedded API call:

## Steps

1. Run: curl -H “Authorization: token $GITHUB_TOKEN” \

https://api.github.com/notifications

2. Parse the JSON response

3. Return action-required items

This breaks when GitHub changes its API version. It fails silently if the token isn’t exported. It has no error handling for rate limits. It doesn’t benefit from the MCP server’s built-in auth handling or retry logic.

Correct — skill that instructs use of MCP tools:

## Steps

1. Call github/get_notifications with params: {all: false, participating: false}

2. Filter to reason = “review_requested” OR “mention” OR “assign”

3. Return structured JSON: {action_required: [...], checked_at: “ISO 8601”}

4. If github/get_notifications fails: return {error: “github_api_unavailable”}

The second version delegates the API call entirely to the MCP server. The skill handles the reasoning — what to filter, what to return, how to handle failure. The server handles the execution.

---

Github Link:

https://github.com/sysdr/openclaw-academy/tree/main/part-2/issue-10/vault-files