Two pipes. Same question.
An MCP and a personal agent do not share a pipe. They land on the same page because the question is the same: what ran.
MCP is a host calling a tool on your server. Cursor, Claude Code, the rest. observe() runs on that process and keeps the tool name, how long it took, and the error. Hosts keep calling your public URL. We do not proxy the tool.
A personal agent has no public MCP URL. A person texts. The work is calendar, mail, booking. Agent() wraps the dispatcher. After a job, the wrap looks for a list or get sibling on the machine. You do not set a second URL.
Same dashboard tab. Different function. The wrong one is rejected.
Found, or miss
A write is not proof the work is there. The second look is. Name the write tool, then the read tool. We pass the new id on the same request. If the read comes back with the thing, it is found. If the write ran and the read is empty, it is a miss.
No map means no second call. We keep the job and the time. We do not invent a check.
On the page, each row is a job. Jobs from one session stay together. A miss after a ship is the cut. Email or webhook, the channels you already use.
What we keep
The wrap checks calendar and mail on your machine. We never sign into Google. We keep the job name, how long it took, and whether it found it. We do not keep names, the chat, or calendar contents.
If the wrap misses the call, the tool still runs. We do not block the job.
That page is Live Traffic.
