Human Approval on Andreas Nissen

Runbook Relay WebMCP Demo

Mon, 01 Jan 0001 00:00:00 +0000

The problem

Incident dashboards are dense, stateful, and consequential. Screenshot-driven automation has to infer what a chart means, which change is selected, and whether an operator approved execution. That is too much ambiguity for an operational control surface.

What I built

Runbook Relay exposes narrow WebMCP tools for reading an incident, comparing mitigations, staging a change, executing an approved change, and resetting the simulation. The tools and the human interface use the same state transitions.

7DayFocus AI Delivery Lab

Mon, 01 Jan 0001 00:00:00 +0000

The problem

An AI planner should not silently rewrite a person’s week. Model output can be malformed, stale, or incompatible with real scheduling limits, and an API key should not become application data.

What I built

7DayFocus is a local-first weekly planner with a proposal-only AI assistant. A user can select Anthropic, OpenAI, or OpenRouter, supply a key for one request, and ask the model to rebalance existing tasks. The assistant cannot create, rewrite, delete, complete, or automatically apply tasks.

From Screenshots to Governed Tools

Sat, 29 Aug 2026 00:00:00 +0000

An agent can use a website by looking at pixels, clicking controls, and reading the result. That is powerful because it works without changing the application.

It is also the least explicit way to automate a consequential interface.

The agent has to infer what a chart means, which control maps to which action, whether the page changed, and whether an approval actually exists. The human sees a visual interface. The agent reconstructs a hidden API from appearance and behavior.