A real MCP session: the agent discovers the suite via .well-known/mcp.json, selects a tool, constructs a typed invocation, and receives a deterministic JSON response. Zero server calls. Zero inference. Client-side replay.
Five discrete steps: no magic, no inference. Here is exactly what each phase does.
.well-known/mcp.json at the root. It points to mcp/server.json for server capabilities and mcp/catalog.json for the full tool list with typed schemas. No API key needed: both are public and machine-readable.manifests/{tool}.manifest.json, and each catalog entry carries the same inputSchema inline (JSON Schema Draft-07). The agent constructs a validated call: no prompt engineering, no natural-language parsing required.These tools appear in real agent invocation demos. Open any one and interact directly: client-side, zero PII.
Point your MCP-compatible agent at the AINumbers server endpoint. The registry loads once; tool schemas resolve on demand.
Connect once. Discover the tool set via .well-known/mcp.json, which resolves to mcp/catalog.json. Each tool's manifest under manifests/ provides the full inputSchema, outputSchema, and policy_mandate envelope spec.
{
"schema_version": "well-known-mcp-v1",
"servers": [
{
"id": "ainumbers-fintech-suite",
"server_url": "https://ainumbers.co/mcp/server.json",
"catalog_url": "https://ainumbers.co/mcp/catalog.json",
"execution_model": "browser-reference"
},
{
"id": "ainumbers-apps",
"endpoint_url": "https://mcp.ainumbers.co/mcp",
"execution_model": "mcp-apps-widgets"
}
],
"llms_txt": "https://ainumbers.co/llms.txt",
"sitemap": "https://ainumbers.co/sitemap.xml"
}
{
"name": "simulate_fraud_score",
"inputSchema": {
"type": "object",
"required": ["amount"]
},
"metadata": {
"tool_id": "fraud-score-simulator",
"category": "fraud-risk",
"url": "https://ainumbers.co/tools/04-fraud-score-simulator.html",
"ap2_export": true,
"execution_type": "browser-reference"
}
}