The weights in the matcher are chosen by hand, and two of them were
adjusted last week because they produced a wrong answer on real data. A
pair agreeing to within 1%, from a merchant the matcher recognised, still
scored 0.62 because a date had drifted. That says more about the
constants than about the receipt.
So the formula keeps what it is good at and hands over what it is not.
Above 0.80 it decides alone: an exact amount from a recognised merchant
needs no second opinion, and buying one for every pair would be latency
and cost for a verdict nobody doubts. Below 0.60 nothing is asked either,
because inviting a model to rescue a pair the evidence does not support
is how plausible wrong answers get made. Only the band between goes for
adjudication, which on a real ledger was five pairs against one the
formula had already settled.
The question is a yes or no with a reason, never a score. An earlier
design in this feature asked a model to rate its own certainty and it
anchored on round numbers, which is what the calibration literature
predicts. Judging concrete evidence and explaining the judgement is a
different task.
A verdict is checked, not trusted: a key nobody asked about is dropped, a
pair answered twice keeps its first answer, and a failed call accepts
nothing, leaving the run exactly where the arithmetic left it. The
formula's own score is stored unflattered next to the verdict, because
dressing it up would hide the uncertainty that sent the pair for a second
opinion in the first place, and agent_metadata records which instrument
decided.
Adjudication runs on a dry run too. A provkörning is meant to show what a
real run would propose, and skipping it would show a smaller, different
answer than the one that lands.
Measured on a real ledger: five uncertain pairs asked, two accepted with
reasons a human can check ("Samma leverantör och belopp stämmer inom
rimlig valutamarginal"), three rejected. Proposals went from one to three.
Co-authored-by: Jakob Wennberg <311770904+jakobwennberg-oss@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Accounted
Open-source Swedish accounting software for sole traders (enskild firma) and limited companies (aktiebolag). Double-entry bookkeeping that complies with Swedish accounting law, built to be operated by you or by your AI agent.
Website · Hosted app · Documentation
Why Accounted?
Compliant by construction. Accounted implements double-entry bookkeeping under Swedish accounting law (Bokföringslagen). Voucher immutability, sequential voucher numbering, period locks, and 7-year document retention are enforced by database triggers, not by convention. Corrections are made the legal way, with reversal entries (storno), never by editing history. See ARCHITECTURE.md for how.
Agent-native. The full bookkeeping engine is exposed as 100+ MCP (Model Context Protocol) tools with scoped API keys, so an AI agent can do the books in Accounted: categorize transactions, draft vouchers, reconcile periods, and prepare declarations. Posting is staged for human approval, so the agent proposes and you decide.
Yours to run. AGPL-3.0 licensed and fully self-hostable with Docker and Supabase. Use the hosted version at app.gnubok.se or run your own.
Features
- Double-entry bookkeeping -- BAS 2026 chart of accounts, draft/commit workflow, sequential voucher numbering
- Invoicing -- Create, send, and track invoices with mixed VAT rates and PDF generation
- Bank reconciliation -- PSD2 bank connection via Enable Banking, 4-pass automatic matching
- VAT declaration -- SKV 4700 form mapping, per-rate breakdown, EU/export handling
- Tax reports -- NE-bilaga, INK2, SRU export for Skatteverket
- Payroll -- Salary runs, payslips, and AGI (arbetsgivardeklaration) employer declarations
- Supplier invoices -- Registration, payment tracking, input VAT deduction
- Document archive -- SHA-256 integrity, 7-year retention enforcement, full archive ZIP export
- SIE import/export -- Standard Swedish accounting interchange format
- Agent access (MCP) -- 100+ bookkeeping tools over the Model Context Protocol, with scoped API keys and staged approvals
- Extension system -- Opt-in plugins for AI categorization, receipt OCR, email, calendar, and more
Self-Hosting
git clone https://github.com/erp-mafia/accounted.git
cd accounted
./setup.sh # Prompts for Supabase credentials, generates .env
docker compose up -d
You need a Supabase project and must apply the database migrations before first use. See docs/SELF-HOSTING.md for the full step-by-step guide, including Supabase setup, auth configuration, optional features (AI, email, push notifications), and troubleshooting.
Development Setup
Prerequisites: Node.js 20+, a Supabase project.
npm install
npm run dev # Start dev server (auto-generates extension registry)
npm test # Run tests
npm run build # Production build
npm run lint # ESLint
See CONTRIBUTING.md for the full development workflow.
Tech Stack
- Framework: Next.js 16 (App Router), React 19, TypeScript (strict)
- Database: Supabase (PostgreSQL + Row Level Security + email/password auth + TOTP MFA)
- Styling: Tailwind CSS 4 + shadcn/ui
- Integrations: Enable Banking (PSD2), Anthropic SDK, LangChain, OpenAI, Resend, JSZip
Documentation
- User and API documentation -- Guides for using the product and the public API
- ARCHITECTURE.md -- How the system is built: bookkeeping engine, legal enforcement, tenancy, extensions, agent surface
- docs/SELF-HOSTING.md -- Full self-hosting guide (Docker, Supabase setup, migrations, optional features)
- docs/EXTENSIONS.md -- Extension development guide
- docs/DOCKER.md -- Docker deployment reference
- CONTRIBUTING.md -- Development workflow, code style, pull request process
- SECURITY.md -- Vulnerability reporting policy
Community
- Found a bug or have an idea? Open an issue
- Security vulnerabilities: see SECURITY.md, never a public issue
- Everyone interacting in the project is expected to follow the Code of Conduct
Contributing
Contributions are welcome. See CONTRIBUTING.md for the full guide.
All commits require a DCO sign-off (git commit -s).
License
AGPL-3.0-or-later with an extension exception: third-party extensions that interact solely through the documented Extension API may be licensed under any terms, including proprietary. See LICENSE for details and NOTICE for third-party attributions.