Track-aware routing
backend, ui, and fullstack stories take different paths — UX design and visual QA only where they belong.
$ /ship-prd
● STORY-001 track:backend
│
├─ PM project-manager READY_FOR_ARCH
├─ ARCH solution-architect READY_FOR_LLD
├─ TL tech-lead READY_FOR_BUILD
├─ DEV developer READY_FOR_QA
└─ QA qa-engineer DONE
● STORY-002 track:ui
│
├─ UX ux-designer READY_FOR_BUILD
├─ FE frontend-developer READY_FOR_QA
├─ QA qa-engineer READY_FOR_VISUAL_QA
└─ VQA visual-qa DONE
Queue score 2/2 stories shipped passed
Evaluate
Most agent setups dump everything into one context window. fullstack-swarm keeps a typed status machine in the main session and dispatches one specialized subagent per stage.
Subagents never call each other — every handoff goes through the orchestrator — so you get clear artifacts and a queue you can inspect with /status.
backend, ui, and fullstack stories take different paths — UX design and visual QA only where they belong.
Every story lives in stories/_queue.json with a status, deps, and history. No black-box loops.
UI stories don’t ship on green unit tests alone — a browser-driven visual check is the final gate to DONE.
Flexible techniques
Compose a full delivery loop from a single PRD — plain-English requirements in, scored artifacts out.
The swarm
Each agent runs in its own context with restricted tools, scoped to one job.
| Subagent | Tracks | Job |
|---|---|---|
project-manager | all | PRD → atomic, track-tagged stories |
solution-architect | all | HLD + ADRs |
tech-lead | all | LLD + confirmed track |
ux-designer | ui, fullstack | UX spec + design tokens |
developer | backend | Implement + tests under src/ |
frontend-developer | ui, fullstack | Implement + tests under web/ |
qa-engineer | all | Functional QA report |
visual-qa | ui, fullstack | Live browser checks + screenshots |
Trace, grade, iterate
Works as a plugin for Cursor and Claude Code. On Claude, add the GitHub repo as a marketplace, then install by plugin name. Scaffold once, write a PRD, then drive the pipeline with slash commands.
git clone git@github.com:mandarzope/fullstack-swarm.git
mkdir -p ~/.cursor/plugins/local
ln -s "$(pwd)/fullstack-swarm" ~/.cursor/plugins/local/fullstack-swarm
claude plugin marketplace add mandarzope/fullstack-swarm
claude plugin install fullstack-swarm
# Use the plugin name fullstack-swarm,
# not the repo path.
/swarm-init
# write prd/current.md
/ship-prd
/status
/next
Commands
/swarm-initScaffold dirs, agents, and orchestrator rules into the target repo.
/ship-prdDecompose prd/current.md and run the autonomous pipeline.
/statusRead-only queue table plus the exact next dispatch.
/nextKeep dispatching eligible stories until DONE, BLOCKED, or stuck.
Any editor. Any track.
Plug fullstack-swarm into the tools you already ship with — no rewrite required.
Ah yes, FAQs.
An eight-agent software delivery swarm packaged as a Cursor / Claude Code plugin. It turns a PRD into working backend and UI code through a typed status machine and specialized subagents.
Subagents cannot dispatch each other. Every handoff routes through the orchestrator, producing inspectable artifacts (stories/, design/, src/, web/, tests/) instead of an opaque chat transcript.
Add the GitHub repo as a marketplace, then install by plugin name:
claude plugin marketplace add mandarzope/fullstack-swarm
claude plugin install fullstack-swarm
Use fullstack-swarm (the plugin name), not the repo path.
No. Install whichever you use. The same agents, commands, and scaffold work in both environments.
backend skips UX and visual QA. ui and fullstack go through UX design and browser-driven visual QA before DONE.
Yes. Run /status to see the queue, then /next to continue dispatching from where the pipeline left off.
Ready when you are
Clone the plugin, run /swarm-init, drop a PRD, and let the swarm take it from there.