When to Use MCP vs CLI
It's not a religion. It's a decision framework.
✅ Use a CLI when...
- Stateless: query in, data out
osgwiki-cli, lvt, git, az kusto
- Local tool on your machine
build tools, debuggers, file analysis
- You want it to work with any agent
Claude Code, Copilot, Codex, Cursor...
- Speed matters (no startup overhead)
✅ Use MCP when...
- Remote service behind enterprise auth
ICM, ADO via Agency proxy
- Auth is the hard part (EntraID, SSO)
agency mcp remote --url ...
- The service already has an MCP server
Slack, Sentry, hosted databases
- Containerized/shared team deployment
The heuristic: Can the agent just run a command and read stdout?
If yes → CLI. If the blocker is auth or remote access → MCP.
“Even if you make a GUI app, add a cli so agents can test the core logic directly — much faster iteration cycles.”
— Peter Steinberger (@steipete), Dec 25, 2025 · https://x.com/steipete/status/2004303530887405889