The ERP that knows things have a history.
The operating system for the circular economy. Built for businesses that sell used, donated, and refurbished goods — refurbishers, Brockenhäuser, repair workshops, vintage shops.
Live demo · Product brief · Docs · Architecture
Every generic ERP assumes you buy new and sell new. A Brockenhaus volunteer accepting a box of donated kitchenware, an IT refurbisher logging fifty laptops from a corporate donor, a repair café tracking spare parts across twenty bikes — none of that fits a "purchase → sell" data model. So secondhand businesses run on spreadsheets and 25-year-old desktop ERPs, and the operational drag keeps them from scaling their impact.
Kivvi handles intake, condition grading, repair workflows, flexible pricing, and impact reporting natively, with the Swiss compliance layer (QR-bills, VAT, KMU Kontenrahmen, CAMT bank import) and an AI command bar that understands secondhand workflows.
Open source. Self-hostable. Built in Switzerland.
- Intake is not purchasing. Donations have no purchase price; the "supplier" is often a one-time individual.
- Every item has a condition. "Like new" and "parts only" route differently and price differently.
- Products flow backwards. Trade-ins, take-backs, and returns are the business, not the exception.
- Testing and repair is part of operations. Multi-step workflows, multiple roles, before an item can sell.
- Pricing is flexible. Richtpreise (guide prices), condition-based pricing, consignment splits, "pick your own."
- Inventory is mixed. 500 identical power cables and one specific Dell laptop with one specific dent — in the same system.
- Impact is the point. Items saved, CO2 avoided, people served — first-class metrics, not afterthoughts.
Full positioning in PRODUCT.md.
| Module | What it does |
|---|---|
| Intake | Accept donations, trade-ins, consignment. Condition grading. Donation receipts. |
| Inventory | Mixed model: bulk quantities and per-item tracking with status (intake → testing → repair → ready_for_sale → sold). |
| Repair workflow | Per-item repair log, parts consumption, labour tracking, condition gates before sale. |
| Sales documents | Unified document model — Angebot, Auftrag, Lieferschein, Rechnung, Gutschrift, Mahnung. Convert by changing one field. |
| Swiss QR-bills | Generated automatically. Legally compliant since 2022. |
| Accounting | 227-account KMU Kontenrahmen, journal entries auto-generated, balance sheet + P&L + VAT report. |
| Banking | CAMT.053/054 import, automatic payment matching via QR reference. |
| Impact reporting | Devices saved, kg diverted from landfill, CO2 avoided. Auto-generated annual Vereinsbericht. |
| AI command bar | Cmd+K. "50 laptops donated by UBS" → intake created. Same domain functions the UI uses. |
| Self-service migration | CSV import with auto-detection (kivitendo profiles included). |
A pnpm monorepo. Domain logic is the source of truth — the UI, the public REST API, and the AI tools all delegate to the same functions in packages/core.
kivvi/
├── apps/
│ └── web/ # Next.js 14 App Router, Server Actions, server components
├── packages/
│ ├── database/ # Drizzle ORM schema — THE source of truth
│ ├── core/ # Domain logic (pure functions, one tx per business op)
│ │ └── src/domain/
│ │ ├── documents.ts # Unified document model (quote → invoice → credit note)
│ │ ├── inventory-items.ts # Per-item lifecycle, condition, repair
│ │ ├── accounting.ts # Journal entries, balance sheet, P&L, VAT
│ │ ├── banking.ts # CAMT import, payment matching
│ │ ├── impact.ts # CO2, devices saved, donation receipts
│ │ └── ...
│ └── ai/ # AI providers + tool registry (Anthropic / OpenAI / Ollama)
└── docs/ # Deployment, self-hosting, audit reports, OpenAPI
Data flow. Server Action → domain function (transactional) → Drizzle → Postgres → revalidatePath → UI. The AI command bar calls the same domain functions through tool definitions, so every AI action is auditable in aiActionAudit.
Full engineering bible in CLAUDE.md.
| Layer | Technology | Why this one |
|---|---|---|
| Framework | Next.js 14 (App Router) | Server Components + Server Actions = fewer moving parts |
| Language | TypeScript (strict) | Types derived from Drizzle schema ($inferSelect) — single source of truth |
| Database | PostgreSQL + Drizzle ORM | Full SQL control, best TS inference, no codegen |
| Money | decimal.js |
0.1 + 0.2 !== 0.3 in IEEE 754; financial software cannot use floats |
| Auth | NextAuth.js v5 | Credentials provider, JWT strategy |
| AI | Anthropic Claude · OpenAI · Ollama (self-hosted) | Configurable per company; default Claude |
| UI | Tailwind + shadcn/ui + Radix | CSS variables as the only design SSOT; semantic tokens everywhere |
| Validation | Zod | Schema = SSOT for validation and types |
| Hosting | Self-hosted (Docker · any Linux VPS) | Runs on our own Hetzner box behind Caddy; portable to any host |
# 1. Clone
git clone https://github.com/g-but/kivvi.git && cd kivvi
# 2. Install
pnpm install
# 3. Start Postgres
docker compose up -d postgres
# 4. Env
cp .env.example .env.local
# edit DATABASE_URL, NEXTAUTH_SECRET, and an AI key
# 5. Schema + seed
pnpm db:push
# 6. Run
pnpm devOpen http://localhost:3000. The onboarding wizard creates your company, seeds the 227-account Swiss KMU Kontenrahmen, and offers a kivitendo CSV import if you're migrating.
DATABASE_URL="postgresql://kivvi:kivvi_dev@localhost:5432/kivvi"
NEXTAUTH_URL="http://localhost:3000"
NEXTAUTH_SECRET="$(openssl rand -base64 32)"
# Pick one AI provider (or none — the app runs without AI)
ANTHROPIC_API_KEY="sk-ant-..." # recommended default
# OPENAI_API_KEY="sk-..."
# OLLAMA_BASE_URL="http://localhost:11434"Companies that can't send data to a third-party LLM can run Kivvi end-to-end on their own hardware with Ollama:
docker compose --profile ai up -d ollama
docker exec -it kivvi-ollama-1 ollama pull qwen2.5:32b
# then set OLLAMA_BASE_URL in .env.localSame AI tools, same UX, zero data egress.
Six ground truths, applied ruthlessly. Full reasoning in CLAUDE.md.
- Software exists to serve humans. Complexity must earn its place.
- State defines behavior — one source of truth. Types derived from schema. Constants centralized.
- Change is constant — design for it. DRY, KISS, YAGNI.
- Automate the mechanical. Human attention is for judgment, not data entry.
- Complexity compounds; simplicity scales. Three similar lines beats premature abstraction.
- Correctness beats speed. Wrong-but-fast loses to right-but-deliberate.
Practical consequences:
- Every multi-table operation wrapped in
db.transaction(). - Every query filters by
companyId(multi-tenant isolation is a security invariant, not a feature). - Every monetary calculation uses
decimal.js, rounded at line-item level (Swiss standard). - Every design token lives in
apps/web/app/globals.css— no hex colors in components.
Shipped (used in production by revamp-it and partner shops):
- ✅ Multi-tenant company model, role-based access, team invitations
- ✅ Unified document model — quote, order, delivery note, invoice, credit note, dunning
- ✅ Intake workflow with condition grading and donation receipts
- ✅ Per-item inventory with repair lifecycle (testing → repair → ready_for_sale → sold/returned)
- ✅ Swiss QR-bill generation, CAMT import, automatic payment matching
- ✅ Accounting: 227-account KMU Kontenrahmen, journal entries, balance sheet + P&L + VAT report
- ✅ Recurring invoices with cron-driven generation
- ✅ Impact reporting (devices saved, CO2 avoided)
- ✅ AI command bar (Anthropic, OpenAI, OpenRouter, Ollama)
- ✅ Public REST API + webhooks for ERP-to-warehouse / ERP-to-shop integrations
- ✅ Kivitendo CSV migration with auto-detection
- ✅ German, French, English UI
Next (open issues track these):
- 🛠 Shop-front for marketplace listings (Ricardo, Tutti integration)
- 🛠 Mobile app for warehouse staff (intake on the floor, no laptop)
- 🛠 Autopilot mode — AI takes routine actions with human-review thresholds
PRs welcome. See CONTRIBUTING.md for the loop.
git checkout -b feat/your-change
pnpm install && pnpm dev
# … make changes …
pnpm lint && pnpm type-check && pnpm test
# open a PR; CI runs lint + type-check + 880+ tests + production buildFound a vulnerability? See SECURITY.md. Do not open a public issue for security reports.
Kivvi is built by RevampIT (Zürich) — a nonprofit IT refurbisher running it in production — and the wider community of secondhand shops that want their software to fit their workflow, not the other way around.
If you run a Brockenhaus, refurbishing workshop, or vintage shop and your ERP is fighting you, get in touch.
MIT — see LICENSE. Use it, fork it, sell it, embed it.
Every item that passes through your hands deserves to be tracked, valued, and given its best possible future.