A solo indie developer just updated a tool that lets you run a coordinated team of AI workers on your own computer, free, with no cloud subscription and a built-in kill switch if they start spending too much.
What's happening
Munder Difflin is a free, MIT-licensed desktop app by independent developer Chaitanya Giri. It wraps AI agent command-line tools (Codex, Grok, Kimi, and free local models via Ollama) into a cooperative team that runs entirely on your machine. One coordinator agent, nicknamed Michael, takes your instructions and delegates to the others. The agents share memory, pass tasks through a built-in mailbox, and keep working after you close your laptop.
Version 0.4.5 landed on August 22, and the headline fix is a big one: cost tracking was silently undercounting spend by more than half after a restart. That matters because one of the most compelling features is the cost circuit breaker, a spending cap you set upfront that pauses all agents if the total hits your limit. For a small shop that cannot afford a surprise API bill at month end, this kind of guardrail is essential.
The whole stack runs locally. Files, context, and prompts stay on your machine. Connect Ollama (also free, also open-source) and you can run the entire thing without paying a single API fee.
Try this this week
- Look at the GitHub repo and scan the README. You will need Node.js 18 and some comfort with a terminal. If that sounds like your wheelhouse, install it. If not, share the link with whoever handles the tech side for your team.
- Start with one agent, one task. Do not build a full office on day one. Pick one repetitive thing (drafting a daily summary, triaging a shared inbox, pulling together yesterday's notes) and assign it to a single agent.
- Set a spending cap before you launch anything. The budget limit is configurable during setup. Start it at $5 for the first week. If it runs out before Friday, that is useful data about how much work the agent is actually doing.
- Review the shared memory after a few days. Agents accumulate context as they work. Check what they have held onto and clear anything sensitive before you hand a task over to the next agent.
The bigger picture
Running a team of AI workers used to mean buying an enterprise platform with a five-figure contract and a six-month rollout. Now it fits in a GitHub repo, runs on a laptop you already own, and ships with spending controls a local business can actually configure. The smallest working thing here is one agent, one task, one clear budget limit. Get that running, see what it does, and only then decide whether you want more.