// ai-toolhub.org — field notes in applied AI
Figuring out which AI tools are missing — by building them.
A working notebook from one engineer in Switzerland. I pick a real gap in how we use AI, build something small and real to probe it, and write down what holds up and what doesn’t.
AI you can approve, that remembers, that talks back — three things I’m working out.
What I'm exploring
Three different questions about using AI well — each one I’m probing by building something real, not by writing a manifesto. They share a workshop and a way of thinking, not a product roadmap.
mcp-approval
runningTrust. Gate what an AI agent actually does — write and delete actions signed by a human with a passkey (WYSIWYS).
Learn more →mcp-knowledge
buildingMemory. Durable, structured memory for agents — documents, records and notes, reached worker-to-worker, off the public internet.
Learn more →voice-coach
open questionVoice. An exploration of spoken interaction — finding where talking genuinely beats typing, and how to build for it.
Learn more →How I go about it
- Build to learn. I'd rather take one thing far enough to find out if it actually holds up than collect ten demos that never get tested in anger.
- Small surface, on purpose. Service bindings instead of public APIs, fail-closed defaults, secrets managed not embedded — fewer ways to be wrong.
- Write it down. Infrastructure as code, decisions and trade-offs noted, so I (and anyone reading) can see why it's built this way.
The question underneath the agent tools: policy isn't approval →
What it runs on
Cloudflare Workers · D1 · R2 · MCP (Model Context Protocol) · WebAuthn / passkeys · OAuth 2.1 + PKCE · Terraform (IaC) · TypeScript
About this site
AI ToolHub is an independent, non-commercial site about practical, safe AI tools — agent approval, durable memory and voice interaction. No product, no sign-up: just notes on what actually works.
About this site →