MCP (Model Context Protocol) and the standardization of LLM tools
MCP is becoming the de facto standard to connect tools with LLMs. Why it matters, how it works, how we're using it.
March 20, 2026 · Lixto Labs Team · 1 min read
The problem MCP solves
Until recently, every agent framework had its own format for tools. Migrating from OpenAI to Anthropic meant rewriting all your integrations. Migrating to LangChain added another proprietary abstraction.
MCP (Model Context Protocol), launched by Anthropic in late 2024, proposes an open standard: any MCP server can be consumed by any compatible client. By 2026, virtually every serious framework supports it.
How it works
An MCP server exposes tools, resources, and prompts via a standard protocol (JSON-RPC over stdio or HTTP). Any LLM-host (Claude Desktop, Cursor, your app) can connect, discover capabilities, and use them without specific code.
Why it changed our work at Lixto Labs
- Reuse: the same MCP server we wrote for Supabase + Claude Desktop runs in production with GPT-5 unchanged.
- Tool marketplace: dozens of open source MCP servers ready to connect Slack, GitHub, Notion, Postgres, Stripe, etc. Like npm for AI tools.
- Composability: chain multiple MCP servers in one agent. E.g., the agent reads an email (Gmail server), classifies it, creates a ticket (Zendesk server) — no glue code.
- Auditability: every MCP tool call is loggable. SOC 2 compliance becomes much easier.
When NOT to use MCP
- Trivial integrations: a single HTTP call doesn't need MCP overhead.
- Ultra-low latency requirements: protocol layer adds ~10-50ms per call.
- Highly sensitive data: requires careful auth, rate limiting, logging.
Looking to 2027
We'd bet that in 12-18 months, writing tools "by hand" will be the exception. MCP (or its successor) will be the standard. If you're designing a tool-using system today, consider MCP from the start.