AI AGENT ARCHITECTURE

MCP vs CLI
for AI Agents

The debate every PM building AI products needs to understand

The Two Approaches

Every AI agent needs to interact with external tools. There are two fundamentally different ways to do this, and choosing wrong can cost you 17x more per operation.

>_ CLI TOOLS
Traditional command-line tools that agents invoke directly. Think gh for GitHub, aws for AWS, jq for JSON. They’ve worked for 50 years and LLMs already know them deeply from training data.
  • Agent runs shell commands directly on your machine
  • Output piped between tools (Unix philosophy)
  • No protocol layer, no server needed
“Like giving someone a well-stocked toolbox”
♥ MCP SERVERS
A standardized protocol that connects AI models to external services through structured APIs. Created by Anthropic, adopted by Claude, VS Code, Cursor and more. Often called “USB-C for LLMs” because any client can connect to any server.
  • Structured API with OAuth, scoping, and audit trails
  • Runs remotely or locally, any platform
  • Universal discovery via connectors directory
“Like a universal adapter with built-in safety checks”
WHY THIS MATTERS NOW
A Hacker News post titled “MCP is dead, long live the CLI” sparked a heated industry debate in early 2026. Perplexity’s CTO publicly moved away from MCP. But the real answer is more nuanced than either side admits.
AI AGENT ARCHITECTURE

How Each One Works

CLI AGENT FLOW
User Prompt
→
Agent Reasons
→
Runs Shell Command
→
Returns Result
Direct execution. The agent runs shell commands like gh pr list --json exactly as a developer would type them in a terminal.
Composable by design. Output from one command pipes into the next naturally. LLMs know Unix pipe patterns deeply from training data.
Zero overhead. No protocol negotiation, no schema injection, no server to configure. If the CLI is installed, the agent uses it immediately.
MCP SERVER FLOW
User Prompt
→
MCP Client
→
Transport Layer
→
OAuth Check
→
API Call
→
Structured Response
Structured and secure. Every request passes through defined protocol layers with OAuth scoping that controls exactly what the agent can access.
Universal compatibility. The same MCP server works with Claude, VS Code, Cursor, and any conforming client. Build once, connect everywhere.
Remote by default. The actual computation happens on the MCP server, not on the user’s machine. Their local environment stays clean and protected.
THE TRADE-OFF IN ONE LINE
CLI is a straight line from prompt to result. MCP adds checkpoints at every step, which costs tokens but buys you security, governance, and universal compatibility.
AI AGENT ARCHITECTURE

The Numbers Don't Lie

Tokens Used Per Task (Claude Sonnet 4)
Task CLI MCP Diff
Repo language check 1,365 44,026 32x
PR review status 1,648 32,279 20x
Repo metadata 9,386 82,835 9x
PRs by contributor 5,010 33,712 7x
Release & deps 8,750 37,402 4x
Source: ScaleKit benchmark, 2026
Reliability & Cost
100%
CLI Reliability
25/25 runs successful
72%
MCP Reliability
18/25 runs (7 TCP timeouts)
17x
Cost Difference
$3.20 vs $55.20/month at 10K ops
ROOT CAUSE
GitHub's MCP server injects all 43 tool definitions into the context window on every call, even when only 1–2 tools are actually needed. That schema alone costs thousands of tokens before any work begins. MCP gateways with schema filtering can reduce this by ~90%, bringing costs down to ~$5/month, but they add another layer of infrastructure.
AI AGENT ARCHITECTURE

When to Use Which

Use CLI When…
Automating your own workflow
You’re a developer building for yourself or your team. You control the environment and know the tools.
Token cost matters
At 10–32x cheaper per operation, CLI is the obvious choice when you’re paying per token at scale.
You need composability
Piping gh pr list | jq is natural for LLMs. They’ve seen millions of pipe chain examples in training data.
Tools have well-known CLIs
GitHub (gh), AWS (aws), Google Cloud (gcloud), Kubernetes (kubectl). The model already knows these deeply.
Speed and debuggability matter
Direct shell execution means instant feedback. You can inspect every input and output at every step.
Technical audience only
Your users are developers comfortable with terminals, sandboxes, and environment setup.
Use MCP When…
1
Building products for customers
When your agent acts on behalf of customers, touching their data in their orgs. CLI can’t scope per-user permissions.
2
Enterprise security is required
OAuth, incremental scope upgrades, token revocation, structured audit trails. Compliance teams need these guarantees.
3
Users are non-technical
You can’t teach designers to use a terminal. MCP servers plug into Claude, VS Code, and other tools they already use.
4
Cross-platform consistency
One MCP server works on Windows, macOS, and Linux without per-platform CLI installation and configuration.
5
You need remote execution
Remote MCP servers keep operations off the user’s machine. No local access risk, no sandbox configuration needed.
6
Distribution matters
The Claude Connectors directory lets users add your service in one click. No installation, no setup, instant access.
“There is no binary choice here. You have to use what gets the job done.”
Den Delimarsky, MCP Core Maintainer
AI Agent Architecture

The PM’s Playbook

What the Smartest Teams Do
1
CLI for development workflows
Use CLI tools when your engineering team is automating their own processes. Git workflows, test running, debugging, repo management. The 10–32x cost savings compound quickly at scale.
2
MCP for customer-facing features
When your product serves customers across organizations, MCP provides the OAuth scoping, audit trails, and governance that enterprise buyers require before signing contracts.
3
Evaluate per integration
Not every tool needs the same approach. GitHub works great as CLI. Figma works better as MCP because designers won’t use a terminal. Evaluate each integration individually.
Quick Decision Matrix
Solo developer?
CLI
B2B SaaS product?
MCP
Technical users?
CLI
Non-technical users?
MCP
Cost-sensitive?
CLI
Compliance required?
MCP
Key Takeaways
  • → CLI is 10–32x cheaper and 100% reliable vs MCP’s 72% in benchmarks
  • → MCP’s value is not speed or cost. It’s security, governance, and universal compatibility.
  • → LLMs already know CLI tools deeply from training data. Zero-shot usage just works.
  • → The “MCP is dead” narrative misses the enterprise use case entirely
“MCP is not just the protocol. It’s a universal standard by which anybody of any skill on any platform can use AI-powered services right away.”
Den Delimarsky, MCP Core Maintainer
Rizvi Haider by Rizvi Haider