Open source · Cloud or self-hosted · Config-driven
Koji turns any document — PDFs, Word, scans, contracts, forms — into clean, structured JSON. Run it on our cloud in two minutes, or self-host the open source core on your own infrastructure. Same product, your call.
$ uv tool install git+https://github.com/getkoji/koji.git
$ koji init myproject --template invoice
$ koji process invoice.pdf
→ output/invoice.json (97% confidence, 2.4s)01 Manifesto
Every contract, invoice, and form in your business contains data worth extracting. But the tools that extract it expect you to hand your data over — to their cloud, their servers, their terms of service.
Koji doesn't work that way. It runs on your infrastructure. It uses the models you choose. It shows you exactly what it's doing. And it's open source, forever.
— The thing we wished existed when we needed it.
02 How it works
No black boxes. Every stage is inspectable, configurable, and replaceable. The pipeline is just a YAML file.
Declare what you want to extract in a few lines of YAML. Field names, types, descriptions. The description guides the model.
name: invoice
fields:
invoice_number:
type: string
required: true
total_amount:
type: number
line_items:
type: arrayPoint Koji at a file or a directory. It parses, classifies, routes, and extracts — locally, in containers you control.
$ koji process ./invoice.pdf \
--schema schemas/invoice.yaml
→ parsing [████████] 1.2s
→ extracting[████████] 2.4s
→ validating[████████] 0.3s
✓ 11 fields, 97% confidenceClean JSON, per-field confidence scores, and the chunks each field came from. Pipe it anywhere.
{
"invoice_number": "INV-2026-0042",
"date": "2026-03-15",
"vendor": "Acme Corp",
"total_amount": 4250.00,
"currency": "USD"
}03 Declarative extraction
Schemas are YAML. Pipelines are YAML. The model sees your field descriptions, not your prose. No prompt engineering, no fragile string templates — just a config file you can version, diff, and share.
name: invoice
description: Commercial invoice extraction
fields:
invoice_number:
type: string
required: true
description: Unique invoice identifier
invoice_date:
type: date
required: true
total_amount:
type: number
required: true
description: Grand total including tax
line_items:
type: array
description: Items purchased
items:
type: object
properties:
description: { type: string }
quantity: { type: number }
amount: { type: number }{
"invoice_number": "INV-2026-0042",
"invoice_date": "2026-03-15",
"total_amount": 4250.00,
"line_items": [
{
"description": "Strategic consulting",
"quantity": 8,
"amount": 1600
},
{
"description": "Documentation package",
"quantity": 1,
"amount": 350
},
{
"description": "Training workshop",
"quantity": 2,
"amount": 1150
}
]
}04 Two ways to run Koji
Most platforms force you to pick a side — hosted forever, or roll your own forever. Koji is built to be both. One product, two operators. Pick the path that fits, swap later if it stops fitting.
We run it. You upload and go.
You run it. Full control, full source.
uv tool install git+https://github.com/getkoji/koji.git, then koji start. Docker Compose under the hood.
One codebase. Same dashboard, same schemas, same pipelines, same CLI. The only difference is who operates the servers.
05 Data sovereignty
Most document processing platforms were designed around a single assumption: that you'll upload your data to their cloud and trust them with it. Koji starts from the opposite premise — and the cloud inherits that assumption, too.
Run it where you need it. Our cloud, your VPC, your laptop, your air-gapped Kubernetes cluster. The architecture is built for the self-hosted case — which means the cloud case inherits real isolation, not a retrofit.
Local models via Ollama, API providers like OpenAI and Anthropic, or your own inference endpoint. Mix and match per pipeline step. Change a line of YAML, not a line of code.
Apache 2.0 licensed. The cloud runs the same code you can read, fork, and self-host. No proprietary extraction logic, no closed model pipeline, no "here's the community edition and here's the real one."
06 Validated
Every change to the extraction pipeline runs against a public corpus of real-world documents — invoices, insurance policies, SEC filings, contracts, forms, receipts. Accuracy is measured, versioned, and published. Regressions get flagged in PRs before they merge.
See full benchmarks →07 Pricing
Pay a flat monthly fee for an allowance of documents. When you need more, overage is a flat per-doc rate. No per-token math, no credit systems, no "platform fees" at the bottom of the invoice.
BYO endpoint All tiers run on your own AI endpoint — OpenAI, Azure OpenAI, Bedrock, Anthropic, Ollama, or any custom model. You pay your model vendor directly; we never touch the inference bill. No markup on tokens we didn't generate.
$0 forever
Evaluate with real documents. 500 docs/month, no credit card.
$499 /month
For teams running production extraction workloads.
Custom
Compliance, long documents, and a dedicated team behind you.
Prefer to self-host? Koji OSS is free forever under the Apache 2.0 license. No limits, no metering, no phone-home. See it on GitHub →
Two ways to start today
Point it at an invoice, a contract, a form — anything you've been meaning to turn into data. Start in the cloud if you want speed, or install the open source core if you want control.
Two minutes from zero to your first extraction. 100 docs/month free forever, $49/month Pro with 500 docs included. Launching soon.
Request early access →Apache 2.0 licensed. Full product, no feature gates. Runs on Docker, docker-compose, or Kubernetes. Your infrastructure, your rules.
$ uv tool install git+https://github.com/getkoji/koji.git