Design an agent for anything, drop it on a one-use microVM, wire it into GitHub or Discord. Runs on your AI subscription. The key never touches the box.
Design an agent for whatever you need — autonomous or interactive, input-driven or output-driven. Pick a model, give it tools, declare its I/O contract. The coding agents are just the first ones we built.
A microVM is a microVM — dough boots one to host a long-lived app or to run an ephemeral agent workload; the compute layer doesn't care which. Nomad-orchestrated, so it scales from one box to a fleet across distributed machines.
Agents run on your ChatGPT/Codex login or your Anthropic key — but the workload only ever holds a disposable, per-task token. The real credential is injected at the network edge and pinned against spoofing. A compromised workload can't leak or borrow anyone's key.
Runs on the subscriptions you already pay for. Codex is your ChatGPT login; Claude is a pay-per-use Anthropic key. Usage bills to you — dough is not a reseller.
One agent's output becomes another's trigger through the Bus. A shared Pad carries state between them for clean hand-offs. Multi-step, multi-agent workflows compose without a human in the loop.
A PR opens, an agent reviews it. You @mention in Discord, an agent replies. Or call the direct API. Integrations are the on-ramp — they embed agents into your day without a new dashboard to babysit.
Per-user isolation and bring-your-own credentials today. Self-hosted / on-prem is on the roadmap once the infrastructure code is cleaned up — a real, acknowledged effort. We won't pretend it ships before it does.