How MCP Interception Works

Guardrly uses the Model Context Protocol (MCP) to intercept API calls at the source — before they reach external platforms.

The MCP Layer

When you install Guardrly, it registers two MCP tools:

make_http_request

Replaces the default HTTP tool. Every HTTP call your Agent makes goes through Guardrly first.

agentwatch_log_reasoning

Allows your Agent to log its decision reasoning before taking high-risk actions. This creates a human-readable audit trail.

Interception Flow

AI Agent
→ make_http_request(url, method, headers, body)
→ Guardrly MCP Server
→ Platform detection (shopify/meta/stripe/generic)
→ PII scrubbing (Authorization headers, tokens, emails)
→ Endpoint normalization (/products/123 → /products/{id})
→ Risk assessment (0–3 scale)
→ Local SQLite queue
→ Original API endpoint (request forwarded unchanged)
→ Response returned to Agent

Background (every 30s):
→ Log shipper sends queue to cloud API
→ Semantic parser labels operations
→ Alert engine evaluates 8 rules
→ Notifications sent if needed

Zero Performance Impact

The interception adds less than 5ms latency. All processing is async. Your Agent never waits for cloud operations.

Supported Platforms

Platform	Detection	Semantic Rules
Shopify Admin API	myshopify.com hostname	50 rules
Meta Ads API	graph.facebook.com	50 rules
Generic HTTP	All other URLs	Basic logging