AINumbers.co · start here · agents

How MCP runs one workflow

MCP is the plug that lets an AI assistant use outside tools. Here an assistant checks whether a web bot really is who it claims to be, using three AINumbers tools over MCP. The middle column shows every message on the wire, in order.

Goal of this animation

Show every message that passes between an AI assistant and the AINumbers MCP server while it runs one three-step workflow, and where each part of the work happens.

🔒 No personal dataThe inputs in each call are processed once on our worker and never stored or logged. Every value shown is synthetic, so send synthetic or anonymised data only.
☁ Our server is Cloudflare“AINumbers server” means Cloudflare today. The MCP worker runs on Cloudflare Workers.
Why every result here is green
● green · the check passed and the step may go ahead● amber · held, flagged for review or waiting on something● red · refused, or a mismatch was caught
ALL GREEN ON PURPOSEThe bot in this example is genuine, so the signature is ACCEPT, the key directory VALID and the agent card TRUSTED.
WHAT A NO WOULD LOOK LIKEA forged signature would make art-129 answer REFUSE, a bad directory INVALID and a mismatched card UNTRUSTED. The refusal still returns over MCP as an artifact with its own execution_hash, and the assistant would stop and tell you why.
THE WIRE DOES NOT CHANGEA no travels in the same tools/call result as a yes. MCP carries the answer and never decides it.
0.0s

Start the walkthrough

Words used here

MCP
Model Context Protocol. A standard way for an AI assistant to find and call tools on a server.
host
The app you chat in, such as Claude. It holds the conversation and sends the tool calls.
tool
One named action with a description and a list of inputs. Each AINumbers tool runs one calculation.
parent hash
The previous step's fingerprint, passed into the next call so the three steps stay linked in order.

Where your inputs go

In this door the calculation runs on our Cloudflare Worker. Your inputs travel there inside each call, are used once and are never stored or logged. The server keeps per-call metadata for abuse monitoring: tool name, a pseudonymous caller hash, success and latency.

To keep inputs on your own device, run the same three pages through WebMCP or by hand. Three doors compares all the options.

Workflow agent-identity-verification from chaingraph.json. The endpoint, tool names, method names, the Mcp-Session-Id header and the one-page tool list come from the MCP worker source. The chat text is an illustrative transcript, the input values are placeholders and the fingerprints are demo SHA-256 values.

←→ back and next · Space play or pause · R restart · all explainers