MindFlock

A private flock of AI agents, started by your ticket queue

Task Management

Developer Tools

Artificial Intelligence

GitHub

Hey Product Hunt 👋 I built MindFlock because I kept losing track of my own AI agents. I usually have three or four coding agents running at once. The bottleneck was never the models — it was me. Six terminal windows, no idea which one actually needed my attention, constantly switching contexts and losing the thread on all of them. And every session started with the same manual ritual: pick a ticket, make a worktree, write a prompt, wait. So I built the thing I wanted. MindFlock polls your tracker — Jira, Linear, GitHub Issues, Shortcut, Asana — for work assigned to you, and turns each ticket into a real session: its own git worktree on its own branch, dependencies installed, an agent already seeded with the ticket's description, acceptance criteria, and comments. Nothing typed. PRs that come back with review comments become sessions the same way. You come in where a human is actually needed: read the diff, then commit → push → PR → merge, one click each. Two things I got wrong along the way, which shaped the whole design: I tried running multiple agents in the same repo first. It breaks in confusing ways — two sessions edit the same file, one reads a stale version of what the other just wrote, and it looks like the model being dumb when it's really your setup. Git worktrees fixed it properly. I tried detecting agent state by reading the terminal output. Unreliable — it misses anything that isn't a visible prompt string and breaks whenever formatting changes. Switching to the agent CLIs' own lifecycle hooks made state deterministic, and let me see things screen-scraping simply can't, like "commit hooks are running right now." The other decision I'd defend hardest: MindFlock never commits, pushes, opens or merges by itself, and never writes to your tracker. Every one of those is your click. A pipeline you can't fully trust is worse than no pipeline. It also runs entirely on your machine — no cloud sandbox, no code uploaded anywhere, your own agent subscription. It's Apache-2.0, free, no accounts. Built solo, evenings and weekends, and I've been using it daily for about a year. I'd love feedback — especially from anyone already running agents in parallel. Where did you hit the ceiling? I've found three or four is where I stop being able to review as fast as they produce, and I'm curious whether that's universal or just me.