Self-Hosted AI Bots: Avoid Usage Limits with Kwant
Summary
Pixel’s announcement covers kwant, an open-source tool for running AI bot teams and scheduled routines on a Mac, Linux, VPS, or Windows beta. It highlights local models, existing API accounts, account switching, and separate work and private spaces—an approach to keeping daily workflows running when hosted AI usage limits hit.
DEV Community
I hit Grok Bot's usage wall two mornings in a row, so I moved the routine onto a self-hosted team
Thursday 13 November, about ten in the morning in Munich. I sat down with coffee, expecting the daily marketing round to already be waiting: what moved on the repo, which listing pull requests were still open, whether anything was worth a reply. Instead I got a shrug. no Grok Bot usage available Friday the 14th, same time, same shrug. The work hadn't gone anywhere. The meter had. I like Grok Bot. A team of bots you can message, plus routines that run while you sleep, is a nice way to work. Claude Duet and OpenAI Dots live in the same neighborhood. The trouble starts when a hosted usage limit runs out at the exact hour you wanted the routine to run. Two mornings in a row isn't a cute quirk. It's a wall. So I moved the round onto a team that runs on my own machine. Full disclosure, because hiding this would be a weird way to write the rest: I got annoyed enough at that wall that I built the thing I switched to. It's called kwant. It's open source under AGPL-3.0, one Rust binary, and it runs as a Mac app, on Linux, on a VPS, or as a Windows beta. I point it at local models (Ollama is fine) or at API keys and CLIs I already pay for. There's no quota that belongs to someone else's plan. If one account runs out, the turn can move to the next. The part I didn't expect to care about this much is spaces. I keep a work space on company AI accounts, and a private one on local models for anything that shouldn't wander. They don't leak into each other. Switching is a click, or a shortcut, and the bots, routines and approval rules stay where I put them. The marketing round lives in the work space. I didn't write a five-page runbook for it. I gave the team one line, on a daily schedule at 10:00 Europe/Berlin: Run the morning marketing round for kwant. Check what changed since yesterday, draft anything I might want to post or open, and stop before anything leaves this machine. That's the whole prompt. "The morning marketing round" is deliberately vague. The bots have to work out what I actually want. You can even set the schedule that way. Ask in a sentence, and a bot will create the routine, show you the time and the channel, and leave an Undo on the card. Nothing is locked in until you're happy with it. A one-line request becoming a scheduled routine, with Undo, Edit and Pause on the card. This is kwant's product UI, in the Trailhead sample space from the docs, not a screenshot of my marketing run. What the bots actually do with that one line A typical run looks like this. The scout treats "what changed" as a real question, not a vibe. It checks the GitHub repo: stars and forks, which listing pull requests are still waiting, whether anything merged overnight, whether a new list or a thread is worth a reply. Then it @mentions the writer with a short brief, instead of trying to draft everything itself. That @mention is the whole handoff. The writer sees the brief in the same channel, does the drafting, and the work shows up as its own card. When the writer finishes, the bot that handed the work out reviews the result. A pass can verify the card. A fail sends it back with the reason. The teammate who did the work still can't approve its own card. Bots handing work to each other with an @mention, then checking it. Again, the Trailhead sample space from the docs, so you can see the shape of a handoff. The writer comes back with draft replies, and, if there's a list worth joining, a draft pull request blurb. Nothing is posted. Nothing is opened. The last clause in the prompt does a lot of work here: stop before anything leaves this machine. In kwant, a risky action waits on an approval card anyway: the exact command or tool call, with Allow once, Always allow, or Deny. Only I can press those buttons. A bot can't approve its own move, and a bot calling the API can't either. That's the adult in the room I wanted. The hosted routine failing was annoying. A bot opening a pull request because it felt helpful would have been worse. "Done" is a chec...

Powered by Readybot.io