MCP for AI Agents
August 23, 2026 · 5 min read
AI agents are powerful reasoners but they're trapped inside a text box. They can't read your files, search your notes, or remember what you told them last week — unless they have tools. MCP gives them those tools through a standard protocol that any server can implement.
The Agent Problem
Without MCP, an AI agent can only:
- • Read what you paste into the chat window
- • Generate text responses
- • Forget everything when the session ends
With MCP, the same agent can search your notes, create files, query databases, send messages, and build on previous context. The agent becomes agentic — it can take actions in the real world, not just talk about them.
How an Agent Uses MCP
1. Connection — Agent connects to the MCP server URL with a Bearer token
2. Discovery — Agent calls tools/list to learn what's available
3. Reasoning — Agent reads tool descriptions to decide which tools solve the user's request
4. Execution — Agent calls tools with appropriate arguments
5. Response — Agent uses tool results to compose a helpful answer
This loop happens transparently. The user says "What did I decide about the auth architecture?" and the agent decides to call search_notes, reads the result, and synthesises an answer. No manual tool selection needed.
Which Agents Support MCP?
| Agent | MCP Support | Config Format |
|---|---|---|
| Claude Desktop | Full support | claude_desktop_config.json |
| Claude Code | Full support | .mcp.json |
| Cursor | Full support | .cursor/mcp.json |
| Kiro | Full support (Powers) | .kiro/settings/mcp.json |
| VS Code (Copilot) | Emerging | Settings UI |
| Cline | Full support | cline_mcp_settings.json |
Real Example: AI Agent + xNotePadAI
Here's what happens when you ask Claude "Save my research on React Server Components":
# Claude decides to call create_note
→ tools/call: create_note
title: "Research: React Server Components"
content: "[the research summary]"
# Server creates note, auto-indexes for search
← "Note created with ID: abc123. Tag: agent-created"
# Next week, you ask: "What did I research about RSC?"
→ tools/call: search_notes
query: "React Server Components research"
← Returns the saved note with content
The Safety Model
Giving an AI agent access to your data sounds risky. MCP addresses this with:
- • Tool annotations — destructive tools are flagged, clients prompt before running them
- • Token scoping — each token accesses only one tenant's data
- • Agent sandboxing — agents can only modify notes they created (tagged
agent-created) - • Rate limiting — 30 req/min prevents runaway loops
- • Instant revocation — revoke a token and access stops immediately
Getting Started
Connect your AI agent to xNotePadAI in under a minute: visit Settings, copy your pre-generated token, and add it to your agent's MCP config. Full setup instructions at /connect/.