For AI agents
One connector for checked business facts
Give your agent official VAT rates, tax-code validity, e-invoice validation and exact arithmetic in one call, with the source and date in every answer.
Try it now without a key: 20 calls a day from one address. For more, get a free key (100 calls a month, no card) on the API keys page.
Why agents use it
Agents spend many tokens searching the web for a tax rate or a code list, and still guess when pages disagree. One call here returns a short answer taken from the official source, with the date we read it and a receipt id you can quote. If something is not covered, the answer says not_covered instead of guessing.
Tested
We gave 60 real-world tasks (written by a separate AI that never saw our tools' answers) to AI agents twice: once on their own (with web search, as agents normally work) and once with this connector. Answers were marked by code against official sources.
| Agent | Right alone | Right with connector | Tokens on code checks | Time on code checks |
|---|---|---|---|---|
| Claude Haiku 5.5 | 48 / 60 | 58 / 60 | 78% fewer | 76% faster |
| Claude Sonnet 5.5 | 55 / 60 | 60 / 60 | 56% fewer | 70% faster |
- Tax and e-invoice code checks (Brazil, Mexico, Poland): the biggest win. Agents used 56-78% fewer tokens and finished 70-76% faster, and got more right (Sonnet 10 to 15 of 15, Haiku 7 to 13 of 15).
- E-invoice validation (EN 16931): more right answers for the smaller model (Haiku 12 to 15 of 15; Sonnet 15 to 15), each failed rule named with a fix. It uses more tokens, because the agent sends the whole invoice.
- VAT rates and business arithmetic: both models already knew these well-known rates and sums (15/15 and 15/15 alone). The connector adds one tool call, so it uses more tokens; what you gain is the official source, the date and a receipt id in every answer.
A non-Claude model without web search (gpt-oss-120b) also improved: 42 of 60 right alone, 55 of 60 with the connector. It fixed out-of-date VAT rates the model remembered.
Added up over all 60 tasks, agents used 44% (Haiku) and 47% (Sonnet) fewer tokens in total, because the code-table tasks cost the most when searched on the web. Per typical task the median token use went UP (+208% Haiku, +271% Sonnet), because for rates and sums the model already knows, a tool call adds a round trip. Use it where the answer must be official and current. Tool replies: p95 37.9 ms.
Tools
| Tool | What it answers | Cost |
|---|---|---|
coverage | What is covered (countries, tables, rulesets, calculation inputs) | Free |
vat_rate | VAT / GST / sales-tax rate on a date, with valid-from/to and official source | 1 call |
code_check | Was a Brazil / Mexico / Poland tax or e-invoice code valid on a date | 1 call per code |
einvoice_check | Official EN 16931 rules on a UBL or CII invoice: valid or which rule fails and how to fix it | 2 calls |
einvoice_rule | Official text of one EN 16931 rule | Free |
verify_calculation | Exact recompute of invoice, GST/VAT, loan, depreciation, FX, inventory, payroll and quote totals | 1 call |
Errors tell the agent what to do next, for example which field is missing or when to retry.
Connect
MCP (Claude, Cursor, VS Code, any MCP client)
{"mcpServers": {"avanciq": {
"type": "http",
"url": "https://api.avanciq.com/connect/mcp",
"headers": {"Authorization": "Bearer YOUR_KEY"}
}}}
Claude Code
claude mcp add --transport http avanciq https://api.avanciq.com/connect/mcp \
--header "Authorization: Bearer YOUR_KEY"
REST and OpenAPI (GPT actions, Copilot Studio, LangChain, n8n, Zapier)
curl -X POST https://api.avanciq.com/connect/v1/tools/vat_rate \
-H "Authorization: Bearer YOUR_KEY" -H "Content-Type: application/json" \
-d '{"country": "EE", "date": "2025-07-01"}'
The full description is in openapi.json (OpenAPI 3.1). Agent-readable summary: llms.txt.
Plans
Same key and plans as the Avanciq API. Calls are shared across all tools.
| Plan | Price a month | Calls |
|---|---|---|
| Free | US$0 | 100 |
| Starter | US$9 | 5,000 |
| Growth | US$29 | 25,000 |
| Scale | US$99 | 150,000 |
| Business | US$499 | 1,000,000 |
Your data
We do not store the invoices, numbers or codes your agent sends. We log only which tool ran, the status and the time it took. Details: Privacy and Terms.
Facts and checks only. Not tax, legal or financial advice.