Spergs, LLC

← All products

dayshift

In progress

One queue for the agents your apps run.

About dayshift

dayshift exists because every app that wanted an agent was launching one itself, and each carried its own copy of the resource limits — so the limits throttled whichever app bothered to implement them while the rest ignored them. Now there is one queue. An app POSTs a job, the broker admits it past the gates, runs it through the configured CLI, and the app collects the answer. Four of Spergs' own apps submit to it today: a portfolio scanner, a book drafter, an object cataloguer and a hardware-fault diagnoser.

Every job is a file on disk, so a restart leaves queued work queued and a run cut off mid-flight is requeued or marked failed rather than lost. A runner is data — its binary, its flags, how it takes the prompt, how to read its answer, and which of its failures mean the job ran out of room rather than the CLI being broken — so adding a new agent CLI is a config entry, not a code change. Conversations resume: a follow-up turn inherits the chat's runner, model, working directory and session, and turns go one at a time. Each app holds its own token and sees only the jobs it submitted; another app's job answers exactly like one that was never there.

A coordinator owns the queue and agents poll out to it, so jobs run across several machines with no inbound port on any of them and no second queue to disagree with the first. An agent takes only work its host can do, a conversation stays pinned to the host holding its session, and health is tracked per runner per host — because a credential expires on one machine and not the others. Stopping an agent drains it rather than killing the jobs it is holding. It comes with a dashboard, a command-line client, and an MCP server so an agent session can drive it.

It is a small C++ service on Boost.Beast built from vendored source, deployed with Ansible and systemd, no Docker. Between hosts it speaks TLS against a CA you run.

What it is not, named here rather than implied: there is no hosted dayshift — the coordinator and its agents run on machines you own — and no licence has been chosen, so it is not offered for download today. Agents belong on a network you trust; they hold a token that can lease work, and that is a LAN or a private mesh, not the open internet. An Android app runs on a handset as of 2026-09-24, signed and delivered by the build service, but it is in no store, so it is not something you can install today either.

At a glance

  • One queue, one set of limits, for every app that submits
  • Jobs are files on disk — a restart loses nothing
  • A new agent CLI is a config entry, not a code change
  • Conversations resume, with the model and directory they started on
  • Each app sees only its own jobs, by token
  • Agents poll out: many hosts, no inbound port, one queue
  • Per-runner, per-host health you can alarm on
  • Dashboard, CLI, and an MCP server an agent session can drive
  • Self-hosted C++ on Ansible + systemd, no Docker

More from Spergs

Other independent software we ship, or are working on.