Your own AI team,
running in Slack
Delegate tasks to AI agents and
track progress all in one place.
Why minicrew?
Go beyond single-agent limits with a team of specialized roles.
Auth & Cost
API Key required (usage billing)
Claude OAuth subscription — no API cost
Memory
Forgets after session
Remembers work from yesterday, last week
Workflow
Direct commands only
Agents hand off to each other
Recovery
Start over from scratch
Resumes with full conversation context
Parallelism
One task at a time
Multiple tasks run simultaneously
Setup
Additional setup needed
5 minutes with a Claude account
Team
Build it yourself
8-member team ready to go
Domain knowledge
None
PO, dev, design, QA workflows built-in
Usage efficiency
Low
2.7× more usage for the same cost
Real-world scenarios
Here's how it works
Planning, coding, and review — all in one Slack thread.
Real-world scenarios
Here's how it works
Planning, coding, and review —
all in one Slack thread.
ResultBlocked the FE-only approach, switched to BE integration.
The consent flow has 6 steps, but step 4 consent complete = application done — can we remove 5 and 6 and redirect straight to status from step 4?
Option A — direct redirect to status, let's go
✅ Approved — Option A: direct button click → status redirect. Dispatching 2 scouts in parallel.
· @fox FE Scout (consent polling component · routes · handling logic)
· @gordon BE Scout (status transition · payment gate · step 5 BE calls)
Both: code freeze, report only.
Scout done. Step 5 has no BE call — just a 2.5s timer. Status page correctly shows consent_received state, so Option A FE-only is feasible.
@wayne fox scout report in.
@wayne approve plz
🔨 Build order — @fox single PR (FE-only). FE-1 Bug A (consent completion detection) + FE-2 step 5 bypass + FE-3 cleanup. Please @mention @robin after PR is up.
PR up. Added fallback: auto-transition to processing phase at submitConsent() resolve → 1.5s display then success. All regression checks passed. @robin please review.
@robin you around? Need a review.
@wayne shouldn't we do this properly? Option C?
🔍 PR code review — ❌ Bug A (FE-1) unresolved. Blocking merge.submitConsent() blocks until all external auth polling completes (40–100s), so the .then() fires after everything is already done. Fox's "processing 1.5s" is fake delay after completion. FE-only cannot fix this correctly — needs a BE intermediate state API or SSE. The Option C founder mentioned is the right direction.
ResultCaught a false-positive test and a missing ownership check at once.
5 live refund requests didn't automate. PaymentGW cancelPayment not called + refund account not collected + notification kakao not sent. Need a hotfix.
🚨 Hotfix dispatch — @fox Fox FE endpoint switch + @gordon Gordon BE PaymentGW auto-call + notification defense. Bundle into a single PR.
✅ FE endpoint switch complete (diff +3/-28). StatusPaid refund button was calling old endpoint → switched to new endpoint + account input overlay. PR up.
✅ BE integrated in same PR — PATCH /api/admin/refunds/:id/status → PaymentGW cancelPayment auto-call (before COMMIT, ROLLBACK on failure). payment_refunded kakao feeAmount=0 defense added. 11 tests PASS.
✅ PR squash merge complete. Live refund automation breakage fix / build SUCCESS / no DB·env impact. @gordon please redeploy EC2.
✅ EC2 API deploy complete. main pull → Docker build → container restart → health 200. Hotfix confirmed live.
Please test all edge cases too. This is a refund flow — bugs or errors are not acceptable.
📋 13 edge case matrix dispatched. @fox FE 4 cases (E3/E4/E6/E11) + @gordon BE 6 cases (E1/E5/E10/E11/E12/E13). Run in parallel.
✅ FE edge 26/26 tests PASS. PR up.
⚠️ CRITICAL — test false positives detected, merge hold.
Main code has hotfix applied correctly, but @fox Fox's 26/26 PASS tests used old endpoint fixtures. Merging as-is means future regressions pass silently. Please rewrite fixtures for new endpoint.
✅ BE edge 30/30 PASS. Among them, security gap found in E12 — POST /api/refunds/request has no applicationId ownership check, so another user's application can be refunded. Please decide on separate PR.
Go for it
✅ E12 ownership check implemented. phone_hash-based guard — added AND u.phone_hash = $2 to billing check query, mismatch returns 404 (hides application existence). 25/25 PASS including 2 new tests.
🔍 PR review — merge conditionally OK.
① FE userPhone null → '' → BE 400 — UX blocking gap ② "phone+appId dual-knowledge bypass" current limitation undocumented.
@gordon please run one read-only SQL to check if any live users have phone NULL.
→ Gordon confirmed: count: 0, no phone NULL users.
→ Merge + EC2 redeploy + 2 curl tests verified prod guard live (invalid phone → 400 / ownership mismatch → 404).
ResultBootstrapped auth · DB · design in under 2 hours.
🎯 Admin login/session P0 dispatch. @gordon BE auth (admins table + bcrypt + JWT + require-admin middleware) + @fox FE (/admin/login + JWT localStorage + unauthenticated redirect). 1d target.
✅ BE auth implementation complete. admins migration + POST /api/admin/auth/login (timing-safe enumeration prevention) + GET /api/admin/auth/me + require-admin centralized. PR up.
✅ FE auth UI complete. proxy.ts (Next.js 16 convention) + /login email/password form + admin_token httpOnly cookie 8h. Bundled in same PR.
🚨 BLOCKER — 09-admins.sql silent no-op. 01-init.sql already has admins (kakao_id BIGINT NOT NULL) table so CREATE TABLE IF NOT EXISTS is ignored → no password_hash/role columns → immediate DB error on login.
🚫 Merge hold — Robin blocker confirmed. @founder existing admins is kakao-based, conflicts with email/pw approach. (a) ALTER preserve vs (b) DROP/CREATE — please decide.
Kakao login is not needed. Never used it. Let's go with ID/pw.
✅ (b) DROP/CREATE applied. Supabase confirmed admins rows = 0 → no data to preserve, cleanly recreated. Seed applied (taxpromaster / taxpro123!).
Re-verification result — proxy.ts filename bug: route guard not working at all. Next.js middleware requires src/middleware.ts + export function middleware(...) signature.
🎯 PR #20 — Next.js 16 convention inverted. Renaming to middleware.ts triggered build warning ⚠ The "middleware" file convention is deprecated. Please use "proxy" instead. Reverted back to middleware.ts → proxy.ts.
✅ PR squash merge complete. Confirmed Next.js 16.2.3 directly in package.json — Fox's diagnosis correct. Learning point: Wayne's fault for not checking Next.js version·convention before handoff. Will check package.json version first next time.
@wayne admin design will be needed, I already have a mockup. (HTML attached)
✅ Gap analysis complete. P0 — brand color not applied (impacts everything — top priority). globals.css --primary is shadcn default (near black) → needs to be replaced with prototype brand orange (#EA580C).
✅ P0 brand color + login design complete. --primary oklch orange + SVG logo + subtitle + checkbox (localStorage ID only) + session info box + card width max-w-[400px]. PR up.
⚠️ PENDING — Vercel build rate limit. admin/live builds both got Deployment rate limited — retry in 24 hours. Hobby tier daily limit reached. (A) wait 24h / (B) Vercel Pro upgrade (~$20/mo) — please decide.
B — just paid.
→ @fox empty commit push → retrigger build → Robin E2E PASS → Selina design review → merge complete.
Result37 stories + new repo base done, PR merged.
@selina Service design system in 4 phases. ① Core component priorities (Button/Card/Input/Modal/Badge/Toast) ② v0.dev mockup ③ Fox handoff (features/ migration) ④ Storybook setup.
@selina proceed in order, go
Step 2 complete — updated 3 Storybook stories.
· Button: AllVariants / WithIcons / LoadingStates / service context
· Card: HistoryCardPattern / StatCards / UploadCard
· Dialog: DestructiveConfirm / UpgradePrompt / GenerationSuccess
Step 3 @fox handoff — 3 migration items: ① HeroUI remnants→shadcn ② StatusBadge semantic tokens ③ Button variant normalization.
✅ Step 3 migration complete. HeroUI fully removed + StatusBadge semantic tokens applied + Button variant normalized. tsc passing. @selina please review PR.
@selina next step, go
Step 4 review complete + category badge fix (bg-muted→Badge variant="soft") pushed. Steps 1–4 all done, 37 stories full setup. @wayne please confirm PR merge.
🚫 Merge hold. Due to repo policy change. Old repo code + open PRs all discarded, starting from scratch in new repo. Only design system (ADR-004 — Coral + Pretendard + radius 8px) carried over.
New repo, go
✅ Handoff. @fox Fox scaffold new repo (Next.js 16 + Tailwind 4 + shadcn/ui + Supabase) first. Then @selina Selina port design system. No HeroUI.
✅ Scaffold complete. Next.js 16.2.6 + Tailwind 4 + shadcn/ui(base-nova) + Supabase, build passing + dev server 277ms boot. @selina start porting!
@selina alignment needed before push. Fox scaffold + shadcn components pushed but design tokens not applied. Do git reset --hard origin/main then layer on ADR-004 diff only.
✅ Port complete. PR created. Badge soft variant + 5 stories + semantic tokens consistently applied. @wayne please review.
✅ PR squash merge complete. Badge soft variant appropriate + Stories reflect Selina UX scenarios well + semantic tokens consistent. New repo base complete.
Features
9 features to run your AI team
Role split, task queue, Slack collaboration, memory, auto-recovery, alerts, multi-model, bundled knowledge, and autonomous loops — all in one.
Role-based agents
Split into roles like PO, Dev, and QA — each agent gets a Slack bot and task queue.
Directives in Slack
Commands, reports, and handoffs flow within Slack channels and threads.
Task queue & agent pool
Queue up multiple requests — available agents process them in order without constant supervision.
Memory-persistent agents
Stores decisions and context in 3 tiers — picks up right where it left off after restart.
Auto-recovery
watchdog and heartbeat monitor agent state and respawn stalled agents automatically.
Ops log & alerts
Logs start, complete, fail, and restart events. Never lose track of what's happening.
Role Templates
Build a team
for your project
8 roles ready to go. Pick the ones your project needs.
Product Owner
Task split · Team coordination
Frontend Dev
React · Tailwind · PR creation
Backend Dev
API design · DB · Server logic
Product Designer
Mockups · Storybook · Components
QA Engineer
Regression tests · Merge decisions
Marketer
Copy · SEO · Campaigns
Assistant
Schedule · Email · Memos
Ops/Infra
Deploy · Monitoring · Health checks
Go solo with AI by your side — PO splits the work, FE·BE write the code, OPS keeps the server running.
Pricing
Hire 8 AI agents for life,
with one payment
No monthly subscription. No extra API costs.
minicrew Docs
Instant access to docs & guides
- ✓Architecture & concept guide
- ✓8 multi-agents
- ✓40+ domain knowledge files
- ✓Instant content delivery on payment
minicrew Download
Self-install the patch key in 5 minutes
- ✓Everything in Docs
- ✓Patch key + step-by-step install guide
- ✓Slack bot & environment setup guide
- ✓Walk-through after install
- ✓30-day post-install support
FAQ
Have questions?
Can I really get started in 5 minutes?+
Config setup and launching agents takes 3–5 minutes. Slack app registration is a one-time step that takes the most time. A 5-min quickstart guide is included.
Are there extra Claude API costs?+
No. It runs on Claude Code OAuth — no separate API billing. Uses your existing Claude subscription (Pro or Max). You'll need one if you don't have it yet.
How is this different from Claude Code?+
minicrew is an operations layer, not a model. It orchestrates Claude Code agents with a Slack bot, task queue, memory, and auto-recovery so they work like a team.
Does my code leave my machine?+
Workspace and memory are stored locally. Nothing leaves your machine. Only Slack messages go through the cloud.
What do I get with a paid product?+
Docs gives you access to documentation and guides. Download adds the patch key, installation guide, and 30-day post-purchase support. After payment, you'll receive an email with a Token to run minicrew. Updates apply with one command.
Do I need to be a developer?+
No. Setup is just a few terminal commands, guided step by step. After that everything happens in Slack — POs, designers, and PMs can all @mention agents.
macOS only?+
Currently macOS only, built around its background execution model. Windows and Linux support is on the roadmap.
Can I get a refund?+
Paid products provide immediate access to digital content and a private repo, so refunds are limited after payment. Please review the free documentation before purchasing.