Den Delimarsky, an MCP core maintainer at Anthropic, breaks down when to use MCP servers versus CLI tools for LLM integrations. Neither is universally better — CLIs excel at token efficiency, composability, and developer workflows, while MCP wins on cross-platform distribution, enterprise security (OAuth, scoped access), and non-technical user accessibility.
Delimarsky references a document by David Soria Parra (MCP co-inventor) called "Design Principles for MCP" on modelcontextprotocol.io. The protocol is built around eight core principles that prioritize adoption, stability, and interoperability over engineering purity or rapid feature additions.
| # | Principle | Meaning |
|---|---|---|
| 1 | Convergence over choice | Opinionated about transports, auth, conventions |
| 2 | Composability over specificity | Predefined primitives (tools, resources, roots, tasks) like Lego bricks |
| 3 | Interop focus | Designed for many clients (Claude, VS Code, Cursor, etc.) |
| 4 | Stability over velocity | Universal compatibility before adding features |
| 5 | Capability over compensation | Core primitives, not patches for model shortcomings |
| 6 | Demonstration over deliberation | Real customer scenarios, not abstract engineering desires |
| 7 | Pragmatism over purity | Trade-offs in service of adoption |
| 8 | Standardization over innovation | Codify proven patterns, don't invent new paradigms |
CLI tools go back 50+ years to Unix terminals. Tools like grep, jq, GitHub CLI, AWS CLI, and Wrangler for Cloudflare work well and compose naturally through pipes. They're token-efficient (no bloated descriptions eating context windows), and Claude Code excels at chaining them together — extracting relevant output and passing it to the next command.
"CLIs are good. They're great. They've worked for 50 years. We have proof."
The pre-MCP alternative wasn't just CLIs — people proposed OpenAPI specs for LLM-accessible APIs. The problem: total variability in auth, output formats, and API structures. MCP solves this with opinionated standards.
Authentication & Authorization: MCP mandates OAuth. Every server follows the same auth conventions. Enterprise customers get incremental scope upgrades, token revocation, and fine-grained access control. CLI tools are often all-or-nothing — you log in with your full account and can do anything your account has access to.
Security & Isolation: CLI tools run on your machine with broad access. Remote MCP servers keep execution isolated — they're APIs that guide the model without exposing your local environment.
Distribution & Accessibility: MCP servers in directories like Claude Connectors let anyone add integrations by copying a URL. No need to install software or learn terminal commands. The "USB-C for LLMs" analogy holds — universal standard across platforms.
"I am not going to teach my grandma how to use a CLI to generate word documents. I would rather have an MCP server for that."
| Use CLIs When | Use MCP Servers When |
|---|---|
| You're a developer comfortable with terminals | Cross-platform distribution matters |
| Token efficiency is critical | Enterprise security guardrails are required |
| You can sandbox execution appropriately | Users are non-technical |
| Claude Code handles the chaining | Remote execution isolation is needed |
The lines between MCP and CLIs are blurring. Agent skills now combine MCP servers, CLIs, and custom scripting. MCP's known issues (verbose tool descriptions bloating context) are being addressed: VS Code added dynamic tool discovery, Cloudflare released code mode, and Anthropic published approaches for abstracting MCP tools into direct code calls.
"There is no binary choice here. There's no right answer. You have to use what gets the job done for you."
"MCP is a very opinionated set of guard rails for APIs."
"I prefer remote MCP servers over local MCP servers because local MCP servers are basically just applications that run in your box."
"It was the analogy that was used for it was that it is the USB-C for LLMs."
"There is no binary choice here. There's no right answer. You have to use what gets the job done for you."
From Den Delimarsky's channel. Delimarsky is a core maintainer of the Model Context Protocol and employed by Anthropic. He provides an insider's perspective on the MCP vs. CLI debate, emphasizing that both tools serve different purposes and will coexist. Particularly valuable for understanding MCP's design philosophy and the enterprise security advantages it offers over CLI-based integrations.