A local AI admin worker for accounting firms
Accountants were the first audience OpenClerq was built for: people who need AI's help with explanations and never want its help with arithmetic.
Short answer
OpenClerq gives an accounting practice AI assistance that keeps client data on its own computers and keeps language models away from the figures. VAT, depreciation and reconciliation formulas run in a deterministic engine that returns a proof record; a local or cloud model explains the result in plain language; scheduled runs handle the recurring checks; and every run is recorded.
Why accountants are careful with AI
- Confidentiality. Client records are covered by professional duties and, often, by engagement letters that restrict where data may go.
- Arithmetic. A language model can produce a plausible wrong figure. In a practice, one is enough. See why AI gets numbers wrong.
- Evidence. If a figure is questioned, you need to show how it was produced, not that a chatbot said so.
OpenClerq is designed around all three: local by default, deterministic for numbers, recorded for every run.
What OpenClerq does for a practice
| Task | How | Status |
|---|---|---|
| VAT in and out of gross amounts, rounded to your policy | Formula spec with a proof record | Today |
| Depreciation schedules | Formula spec; the model explains the method to the client | Today |
| Reconciliation differences | The engine computes the difference; the model lists likely causes to check | Today |
| Explaining a rule or a result in plain language | Ask panel, with a local model for client data | Today |
| Month-end and deadline checklists | Scheduled triggers, each run recorded | Today |
| A record of which model produced which explanation, at what cost | Run history with the steps, model, tokens and cost of every call | Today |
| Checking an exported trial balance or ledger file directly | The tool-calling loop reads the export from an allowed folder | 0.6 |
| Approval before anything is filed or sent | Inline approvals on tool calls | 0.6 |
| Country-specific VAT and payroll rules, and your accounting software's exports | A module built for your jurisdiction and stack | Module |
Example formulas, checked by the engine
Taking 25% VAT out of a gross amount of 99.99, rounded to cents on both lines:
{
"spec": {
"formulas": {
"net": "round(gross / (1 + rate) * 100) / 100",
"vat": "round((gross - net) * 100) / 100"
},
"output_names": ["net", "vat"]
},
"inputs": { "gross": 99.99, "rate": 0.25 }
}The engine returns a net of 79.99 and VAT of 20.00, and a proof hash for that result. Rounding is explicit, so per-line and per-invoice rounding are two different formulas rather than a choice the software makes for you.
| Rule | Formula | Example |
|---|---|---|
| Straight-line depreciation | annual = (cost - residual) / yearsmonthly = round(annual / 12 * 100) / 100 | 12,000 cost, 2,000 residual, 5 years → 2,000 a year, 166.67 a month |
| Reconciliation difference | difference = round((bank_total - ledger_total) * 100) / 100 | 18,437.30 against 18,452.30 → −15.00 |
| Adding VAT | gross = round(net * (1 + rate) * 100) / 100 | Whatever your rates and rounding policy are |
OpenClerq ships no country's tax rules. You write the formulas for your jurisdiction, or we build and maintain them as a module.
What stays with the accountant
- Every filing, submission and payment.
- Judgement on classification and treatment — a suggested code is a suggestion.
- Advice to the client.
- Checking the formulas themselves: the engine is exact, so a wrong formula is exactly wrong every time.
Client data stays in the practice
Run a local model through Ollama or LM Studio and client data never leaves the computer. Where your rules allow a cloud model for some work, the context preview shows exactly what would be sent, and the run history records which provider and model received it. OpenClerq's own files are readable only by the user who runs it, and the gateway accepts requests only from that machine, with a token.
Frequently asked questions
Does OpenClerq integrate with our accounting software?
Not out of the box. It works from figures you give it, and from plain-text exports in a folder you choose once release 0.6 lets the model read them. Integration with a specific package's exports is something we build as a module through Clerq for Business.
Does OpenClerq know our country's VAT rules?
No — the open-source core deliberately ships no jurisdiction-specific rules. You write the formulas, which keeps them visible and reviewable, or commission a module that includes them.
Can a client or auditor see how a figure was produced?
Each engine result comes with a proof record: a calculation id, timestamp, the operation and a hash tied to the inputs and outputs. The run history shows the steps around it and which model explained it.
Does it read PDFs of invoices?
Not yet. OpenClerq currently reads plain-text files, such as CSV exports. PDF and scan reading are not part of release 0.5.