uygnoey Building agents that reach production systems with MCP → Production integration patterns (API / CLI / MCP)
Scope
This skill helps you choose and implement an integration approach so an agent can reach real production systems reliably.
Instructions
- Identify what the agent must do (user intent) and what systems it must access.
- Choose an integration approach:
- Direct API calls for small, single-service integrations.
- CLI automation when a robust CLI exists and shell access is acceptable.
- MCP server when you need reuse across clients/environments and want a dedicated integration layer.
- If using MCP, design the server:
- Group tools around intent (avoid endpoint mirrors).
- Keep tool surface compact.
- For very large surfaces, expose a small “search + execute” interface and orchestrate details in code.
- Add richer semantics only where they improve UX (e.g., interactive UI or structured elicitation).
- Use standardized auth and central token storage for production.
- Optimize client context usage:
- Load tool definitions only when needed.
- Process large tool outputs in code and return summaries.
Bundled resources
references/integration-approaches.mdreferences/mcp-server-design-patterns.mdreferences/context-efficiency.md
Examples
Example: picking an approach
User goal: “Create a ticket from this Slack thread and link it to the incident.”
- If you will support this across Slack, email, web, and IDE: build an MCP server with a single high-level tool such as
create_incident_ticket_from_thread. - If it is only for one internal bot and a few endpoints: direct API calls may be enough.
Example: designing intent-based tools
Bad: jira_create_issue, jira_update_issue, jira_add_comment, … (hundreds of endpoint-like tools)
Good: create_issue_from_thread, triage_issue, summarize_issue_status