Keyboard

Shortcuts

Command palette
K
Home
gh
About
ga
Projects
gp
Expertise
gx
Journal
gj
Contact
gc
Copy email
e
Call
c
Toggle theme
t
Show this help
?

↑↓ to navigate · Enter to open · Esc to close

AI & agents

Model Context Protocol in practice: rewiring tools to LLMs

TL;DR MCP standardizes how an agent calls a tool. The real work: keeping the perimeter (allow-list, sandbox, quota).

22 Sep 2025 · Nicolas Albert

  • mcp
  • agents
  • llm
  • tooling
Model Context Protocol in practice: rewiring tools to LLMs

1 min read Reading 0%

The Model Context Protocol (MCP) solves a real problem: every LLM integration reinvented its own tool-call format. MCP standardizes discovery and execution. The protocol is the easy part. The perimeter is the rest.

What MCP Hub adds

  • A tool registry: each tool has a validated input schema (Zod), a version, and docs.
  • Agents connect over stdio or HTTP and list available capabilities.
  • Every call is logged (who, what, when) for audit.

Guardrails by default

  • Command allow-list: the tool can only call what we declared.
  • Sandbox: isolated execution, no network except whitelist.
  • Quota: calls / window, to bound agent loops.

Without that, « give me a tool » becomes « give me an interpreter ». MCP makes integration clean; guardrails make it safe.

← Back to journal

Contact me Call