Amp launches "boxes": a persistent mini-server with a built-in agent for internal apps, from $18

In one evening you built a bug tracker or a mini-CRM for your team with an agent. It works great, but it has nowhere to live: a laptop is not an option, and you can't be bothered to set up hosting for a tool used by ten people. Amp (a Sourcegraph product) answers this problem with a new entity: boxes.

What it is

A box is a single always-on virtual machine. You can deploy several apps to it, and each gets its own address on onamp.dev. The authors stress that there are no containers or other layers: it is just a computer. The logic is simple: agents are good at working with ordinary computers, so why invent anything else for them.

The workflow looks like this. You do development in "orbs" (the Amp environment) and show intermediate results to colleagues through "portals". When the app is ready, you ask the agent to create a box and deploy to it. Later you can ship new versions from any number of orbs. Access can be closed or require login via Sign In With Amp, and that too is done by asking the agent.

The agent lives next to production

The main idea is not hosting but the fact that every box has its own runner. The agent can be launched directly on the machine, next to the running apps and their data. The announcement gives examples: clean up the team wiki's database, add an entry from a screenshot, analyze data the team has dumped in.

Hence the authors' provocative question: why build a complex admin panel if you can just ask an agent to fix the app's state? For internal tools this is reasonable. An admin panel for three people often costs more than the app itself, while a text request covers 90% of scenarios.

Boxes can be opened to the whole workspace: a team runs internal software on one machine, and every member can ship new versions.

How much it costs

Billing is per minute, and by default a box is always on. The sizes are:

• b1.tiny: 0.5 vCPU (up to 2 on bursts), 2 GB, $18 per 30 days;
• b1.small (default): 1 vCPU (up to 2), 4 GB, $30;
• b1.medium: 2 vCPU, 8 GB, $50;
• b1.large: 4 vCPU, 16 GB, $100.

If you run a box on a schedule from 6:00 to 21:00, prices drop to $11.25, $18.75, $31.25 and $62.50 respectively. That is exactly 37.5% in savings. The prices in the demo video are for the Enterprise plan; personal and team boxes are cheaper.

What to keep in mind

First, this is a vendor announcement, and there are no independent tests. Questions about reliability, fault tolerance and backups are not addressed in the text: one VM remains one VM. Second, the idea of an "agent directly changing production data" is convenient but risky: without snapshots and an action log, one bad request can ruin the database. Third, you are tied to the Amp ecosystem: addresses on onamp.dev, login through Amp, the runner inside. This will not suit a high-traffic product with thousands of users, and the authors do not claim it will.

What to do in practice

If your team already builds internal tools with agents (dashboards, wikis, mini-CRMs), boxes fill the gap between "works on my machine" and "works for everyone". You can start with the cheapest size: ask the agent to create a personal box and deploy one tool there. Before giving the agent access to real data, set up backups yourself and check exactly what it is able to change. If you are building a service for external customers, boxes are not a replacement for regular infrastructure.