SoloSwarm blueprintsSoloSwarm started in mid-2026 as a big idea: an “operating system for one-person companies”, where a swarm of AI agents would run a business for you. Over a few months it took several shapes: a directory of AI tools and workflow playbooks, a SaaS-looking dashboard selling automation blueprints, an AI toolkit for cross-border e-commerce sellers, a general workflow engine, a video localisation pipeline.
None of those shipped to real users. Each one was a platform built before there was a customer, and each stopped at “it works on my machine”.
What we were actually doing
Meanwhile, the operation behind SoloSwarm was running a portfolio of websites with dozens of parallel Claude Code sessions: one coordinator session writes briefs and dispatches sub-sessions, each sub-session works in its own git worktree on one site, and reports back. SEO fixes, redesigns, payment flows, outage triage, research.
That part works. What doesn’t work is watching it:
- With twenty sessions open, you can’t tell at a glance which ones finished, which are stuck, and which are waiting for an answer.
- Sessions running in auto mode can push, deploy, edit DNS or call a payments API, and you find out from the summary afterwards.
- Overnight runs hit usage limits and silently stop.
- The cost per project is invisible until you add it up by hand.
Looking at our own logs made it concrete: nearly every tool call ran under automatic approval, and usage-limit notices showed up dozens of times in a month.
So: mission control
The new SoloSwarm doesn’t try to run your business. It sits next to the agents you already use and does four things:
- A board of every session, with a status that tells you where to look.
- Risk-tiered approvals. Reads and local edits go through. Pushes, deploys, DNS changes, payments and outbound messages wait for a human (on your phone, if you like), and silence means no.
- Cost and budgets per project and client, with a breaker that stops a project’s agents when its budget is spent.
- A timeline per session, so you can see what happened.
It’s open source, runs locally, and starts with Claude Code because that’s what we use every day. Codex CLI, n8n and OpenClaw/Hermes come next.
The blueprints on this site are the other half: the briefs and rules we use, written up so you can copy them.