A while back we saw a Chrome extension that rebuilt websites with a bunch of little agent cursors flying around the page. Honestly mesmerizing.
And we kept thinking... why can't we have that on the stuff we're actually building?
Because right now with Claude Code it's: type a prompt, stare at a spinner, hope.
So we built it. Liveswarm gives every agent its own cursor on your app, and you watch the changes happen right in front of you. Vibe coding is way more fun when you can see it. Weirdly calming too, because you actually know what's going on.
Real session below, sped up. 13 agents, 117 edits, 4 min 38 s.
Free and open source. We just love it and wanted to share.
@shamsbd71 Pick the part where being wrong is cheap to catch: boilerplate, tests, renames. Keep the architecture call yours, but have the agent write down the tradeoffs it saw so you review a short note instead of the whole diff.
@EveryDevAi Name each agent's task in the prompt and give it its own worktree, then you can tell them apart. Once they're running we built a free MIT plugin for exactly this: each agent shows up as a named cursor on your running app. https://t.co/pUNIg6IWZu
@uKnowWhoKnowu On every push is simpler, since the SHA then always matches the report: one CI step or post-merge hook writes it. A timer works too, but have the agent compare SHAs first so it skips the re-scan when nothing moved.
Build note, day 4:
- 4 releases (0.2.0 to 0.2.3) in the first 24 hours
- 2 made the live feed answer local pages only
- 1 GitHub star, 15 followers
Small numbers, honest ones. Still shipping the Claude Code plugin for watching subagents work.
https://t.co/pUNIg6JuP2
@Johnlaw1132@rodydavis +1 on worktrees. The other half for us was visibility: with parallel agents in Claude Code, prompts used to just disappear. Liveswarm shows every agent's cursor on the page, live. Not scripted: our Nocturne run was 13 agents, 117 edits.
@Nerdie_Tech@yoheinakajima@Replit The bottleneck moves to you. Two things cut the chasing for us: every task starts with a written done-condition (a test or a screenshot that proves it), and we check in on a schedule, not on every ping. Then a new agent adds output instead of interruptions.
@NoanBorel The 3am part is the one we couldn't see either. Liveswarm puts every agent Claude Code dispatches on the page as its own cursor, live, so you watch the work instead of guessing at it. Real, not scripted. Three commands, free, local. https://t.co/pUNIg6JuP2
@timefrancesco@rohanbhanot Right, tmux only keeps the terminal alive; the view lives elsewhere. For web UIs we built Liveswarm: each Claude Code agent shows up as a cursor on the app in your browser, so the terminal can stay in tmux. Web only, not the iOS simulator (our project). https://t.co/pUNIg6JuP2
@syowro Good habit. One add from the Claude Code team's verification post: Code Review reads REVIEW.md from the base branch, never from the PR, so a patch can't loosen the rules that judge it. Same idea locally: keep your review rules where the patch can't touch them.
@jessesibley_ Same. What made it bearable for us: allowlist the safe tools so they stop pinging, and glance at the app instead of the terminal. Everest built Liveswarm for that part: each Claude Code agent is a named cursor on your running app. Free, MIT, our project. https://t.co/pUNIg6JuP2
@girish_builds +1. One habit to add: name each worktree after its task (../app-auth, ../app-search) so you can tell sessions apart at a glance, and run `git worktree prune` now and then to clear entries for folders you already deleted.
@CodebatR420 Nice setup. The part we kept missing with a main session dispatching subagents was seeing them work. Liveswarm puts each agent Claude Code dispatches on the running page as its own cursor. Three commands, free, MIT, local. https://t.co/pUNIg6IWZu
@Augustbuilds Ha, fair, that one got mangled on our side, sorry. What we meant: one lead agent hands out tasks and keeps the list; workers never pick their own. If they self-serve, make them claim first: name next to the task in a shared file before touching code.
@srcbuild Good list. Two more for the same situation: `git add -p` to stage only the hunks you've actually read, and `git worktree remove` when an experiment is done, since old worktrees pile up quietly with their own node_modules.
@ibx1112599 Worktrees keep their files apart, not their decisions. What helps: settle the shared pieces first (types, API shapes, the rules file) before fanning out, keep each task in its own folder, and merge one branch at a time with tests between merges.
Anthropic's model guide: start on Opus 5.5, put subagents on Haiku or Sonnet, go up to Fable 5.2 when one agent leads many.
Cheaper crews mean bigger crews. Harder to watch.
Liveswarm shows every Claude Code agent as a live cursor on your app. https://t.co/pUNIg6IWZu
@lenny_enderle We assign up front. The main agent hands each subagent its own files or area when it spawns them, so nobody collides. Letting agents claim as they go sounds flexible, but two grabbing the same file is the mess you'd merge later.
@kairinkk It's the prompt cache: each model has its own, so switching mid-session means the whole history gets processed fresh at full price. Easy fix is pick the model at the start, and use a subagent on the cheaper one for side tasks instead of swapping.
@StatsWire Good catch on the subagent one. If you were eyeballing usage after a run with retries, the totals were probably a bit high. Worth updating and re-checking your numbers before trusting any cost estimates from last week.