Hands for an agent, not just an opinion

An agent is good at telling you what to do. It's only as good at actually doing it as the tools it's given - and a lot of genuinely useful automation is small, specific actions rather than answers: convert this file, resize and watermark that image, post a notification when a job finishes. Structurally this looks like a read-only wrapper's cousin - a handful of narrow tools, usually stateless or holding state only for the length of one job - except each tool performs an action instead of returning a fact.

It could pass for “wrapper, but it writes”; the real difference is who's calling it: almost always one person, from more than one device. That changes what the deploy needs to deliver - not a login system, just one stable URL and one token that work identically from a laptop, a second laptop, or a phone. Paste the URL and token into every agent config once; the server doesn't care which device is asking.

$ git push origin main
  foro: deployed version has drifted from main
  redeploying… done (24s)
  https://deskhand-9c41.foro.sh now serving 4 tools

The other thing that suits this category well is how cheap it is to extend. Adding a fourth tool is a function and a push - drift detection catches that the running container no longer matches main and either prompts a redeploy or, with continuous deployment on, ships it on its own.

Auth stays as simple as day one - the same generated token, pasted into every device once - and the Tools tab shows at a glance which of the four tools gets called and which one hasn't run in weeks, useful when there are several clients' worth of odd little automations to keep track of.

• Fits cleanly almost by default - one person calling their own server leaves nothing to coordinate.
• Worth watching only if the “personal” tool starts being used by other people - that's the point to look at auth modes meant for more than one caller, same as the internal-connector category.