openaccountants/openaccountants

OpenAccountants – AI‑ready tax knowledge library

What it is : A curated, machine‑readable collection of tax “Guides” for 231 jurisdictions. Each Guide is either a source‑cited draft or has been accountant‑reviewed (named professional’s name in the front‑matter). The repo stores the raw guide files (skills/), generated bundles (packages/), an inventory (index.json), and a small Python MCP (Model‑Connector‑Protocol) server that serves the data to LLMs.

Why it matters for AI : LLMs often hallucinate tax numbers. By pointing a model at the OpenAccountants MCP endpoint (https://www.openaccountants.com/api/mcp) the model can retrieve a concrete, cited answer and display the reviewer’s name, turning a vague guess into a verifiable citation.

Key components

• skills/ – human‑authored markdown/JSON guides, the only part meant to be edited directly.
• packages/ – nightly‑generated per‑jurisdiction bundles (JSON, CSV) that the MCP server serves.
• index.json – a machine‑readable catalogue of all guides, their jurisdiction, state and review metadata.
• llms.txt – a list of LLM identifiers that the platform recognises for automatic citation.
• mcp/ – a lightweight Python package (openaccountants-mcp on PyPI) that implements the MCP API used by Claude, ChatGPT, Cursor, etc.
• Docs (REPO‑LAYOUT.md, WEBSITE‑SYNC.md) explain the nightly sync that pulls edits from the web platform into the repo and pushes PR merges back to the live service.

How it works

1. Authoring – accountants create or review a Guide on the OpenAccountants web UI; the system records their GitHub username and writes the file into skills/.
2. Nightly pipeline – a set of bots run:
   • Maths check validates tax band totals and arithmetic.
   • Source watch monitors the original legislation pages for changes.
   • Refresh engine re‑derives stale figures from official sources.
   • Exam runs the latest LLMs against the guides to catch hallucinations.
3. Export – generated bundles are committed to packages/ and the MCP server serves them via a simple HTTP JSON API.
4. Consumption – an LLM calls the MCP endpoint, receives the relevant Guide fragment, and can render it with a citation like “Reviewed by Werner Britz CA(SA)”.

Typical use‑case

User: "What's the combined sales tax rate in Manatee County, Florida for 2026?"
LLM → calls https://www.openaccountants.com/api/mcp with jurisdiction=US-FL-Manatee&year=2026
LLM receives: {rate: 7.5%, source: "fl-sales-tax-guide", reviewed_by: "Jane Doe CPA"}
LLM replies with the figure and the reviewer name.

Who contributes : Professional accountants (named reviewers) and community members who add drafts. Contributions are tracked on GitHub under the author’s username, giving a visible contribution graph.

License : Code under AGPL‑3.0; guide content under a bespoke OpenAccountants Guide License (commercial options also available).

Bottom line : OpenAccountants is a genuine, AI‑focused open‑source project that supplies trustworthy, citation‑ready tax data to large language models, reducing hallucination risk for financial‑related queries.