Model or dataset
openaccountants/openaccountants avatar
openaccountants/openaccountants

OpenAccountants: tax Guides with a named reviewer on the frontmatter

Open-source Tax Guides for AI agents, reviewed by named CPAs/CAs/EAs. 1,000+ Guides across 190+ jurisdictions. MCP server included. Works with Claude, ChatGPT, Cursor.

381 stars58 forksPythonAGPL-3.0

At a glance

What is it?
OpenAccountants publishes tax Guides for AI agents across 231 jurisdictions, split into accountant-reviewed and source-cited draft states, and ships an MCP server so Claude, ChatGPT or Cursor can cite them. The value is the attribution, not the file count.
Who is it for?
Adopt OpenAccountants if you need an AI to cite a tax position that carries a named reviewer and a review date, and you are willing to check the frontmatter before trusting any single Guide. Do not adopt it as a filing engine or as a substitute for a professional opinion on your own facts; the README calls the content general reference, not advice, and warns that Guides may be incomplete, outdated, or wrong.
Can I use it commercially?
Yes, with strict conditions. AGPL-3.0 is a network copyleft licence: if people use a modified version over a network, for example as a hosted service, you must offer them its source code under the same licence.
Is it still maintained?
Yes. The repository last received commits 1 day ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.

DEEP OPEN-SOURCE ANALYSIS

The problem is not tax arithmetic, it is who signs the answer

The README states the premise in one line: every AI can do tax math, none of them can stand behind an answer. That is the gap this repository is built around. A model asked about a combined sales tax rate in Manatee County, Florida will produce a number from training data with no date attached and no one accountable for it. OpenAccountants supplies the missing layer, a Guide file with a named reviewer, a credential and a review date recorded in the frontmatter.

The audience is narrow and worth naming. It is developers building tax features into an assistant, and accountants who want their name attached to the knowledge those assistants read. A freelancer asking what they owe in South Africa is the demonstration case in the README, not the buyer. The buyer is whoever has to answer for the output when it is wrong.

Two Guide states, and the grep that keeps them honest

Every Guide sits in exactly one of two states. Accountant-reviewed means a named, licensed accountant reviewed the complete Guide; the name appears in the frontmatter under reviewed_by and on the public roster in VERIFIERS.md. Source-cited draft means the Guide was written from primary legislation with every figure cited to its source, and has not been professionally reviewed.

That distinction is the most useful design decision in the repository. The README's own stats block, described as live from openaccountants.com and updated by a nightly sync, reports 1,856 Guides across 231 jurisdictions against 176 accountant-reviewed. Read those two numbers together and the reviewed share is under a tenth of the corpus. The repository does not hide this. It puts both figures in the same block and lets you draw the conclusion, which is the honest arrangement and also the one that tells you most Guides are drafts.

How the skills, packages and index fit together

The repository separates authored source from generated output. Contributors edit skills/ only, organised per jurisdiction. Everything else regenerates automatically: packages/ holds per-country bundles, index.json is the machine-readable inventory, and llms-full.txt is a generated LLM entry point. The README is explicit that packages/, index.json and llms-full.txt are generated nightly and must never be edited in a pull request.

The data flow runs from authored Guide to nightly generation to consumption. An MCP client calls the hosted endpoint, the server resolves the question to one or more Guides, and the answer returns with the reviewer attribution attached. The README's worked example loads za-income-tax and za-provisional-tax for a South African freelancer and returns an ITR12 working paper and an IRP6 provisional schedule, closing with the line naming Werner Britz CA(SA) as reviewer. docs/REPO-LAYOUT.md and docs/WEBSITE-SYNC.md carry the architecture and sync detail. If you fork this to run your own pipeline, the generated-artefact boundary is the first thing to respect, because a hand-edited packages/ file will be overwritten at the next sync.

Getting it connected: hosted endpoint or pip install

Two paths exist. The hosted connector is a single URL added to Claude, ChatGPT, Cursor, Windsurf or any MCP client:

https://www.openaccountants.com/api/mcp

The README points to openaccountants.com/connect for guided setup. The self-hosted path is the Python MCP server, installed with pip install openaccountants-mcp, which the README says mirrors the repository's packages/ directory. A third option needs no server at all: download your jurisdiction's folder from packages/ and upload the files to your AI directly, starting with your country's main package and using index.json as the inventory.

There is also a Smithery listing for the server. What the material does not give is the JSON config block for any specific client, so the exact key names depend on the client's MCP setup rather than on anything documented here. The server lives in mcp/ and its PyPI package is openaccountants-mcp, which is what you would pin in a requirements file.

The reviewed count is the limitation, not the jurisdiction count

231 jurisdictions and 1,856 Guides read like coverage. They are not the same as verified coverage. The README's own figures put 176 Guides in the reviewed state, which means the answer a model returns for most jurisdictions is a source-cited draft. Drafts are still cited to primary legislation, which is more than a model's training data offers, but nobody licensed has signed them.

The second limitation is stated by the project itself: Guides may be incomplete, outdated, or wrong for your facts, and the README asks that a qualified professional review outputs before filing, payment, or action. A tax Guide for a jurisdiction is a reference document, not a computation over your situation. It will not know your residency days, your entity structure or your treaty position. Treating a cited Guide as a filed position inverts the intended use.

The third is timing. Releases are dated through 2026, and the stats block is stamped by a nightly sync, so the pipeline is active. Whether any individual Guide reflects the current tax year is a per-file question the frontmatter should answer, and that is a check the reader has to perform rather than one the repository performs for them.

Against a general-purpose model or a commercial tax database

The obvious alternative is asking the model directly. A general model answers instantly, costs nothing extra and needs no MCP configuration. It also has no review date, no named reviewer and no citation to primary legislation, which is exactly the failure the README opens with. The difference is not accuracy on easy questions; it is whether the answer can be traced and attributed when it matters.

The other alternative is a commercial tax research database, the kind sold to accounting firms. Those carry editorial staff and update cycles, but they are built for human search rather than for an agent to call mid-conversation, and they are not open source. OpenAccountants trades editorial depth for machine-readability and public attribution: the Guides are files an MCP server can serve, and the reviewer's name travels with the answer. It also trades control for licensing, which is the next problem.

Licence split, contribution flow and what maintenance actually costs

The licence is split, and the split matters more than the licence name. Code in mcp/, scripts/ and tools/ is AGPL-3.0. Guide content is under the OpenAccountants Guide License v1.0, with commercial options described in COMMERCIAL-LICENSING.md. AGPL-3.0 on the server code means that if you modify the MCP server and expose it over a network, the copyleft obligation attaches to your modified server, not to the Guide content. Anyone planning to embed the Guides in a commercial product should read LICENSING.md and COMMERCIAL-LICENSING.md and take their own advice; the repository provides the documents, not the answer.

Maintenance cost depends on which side you are on. Consuming the hosted endpoint costs nothing to maintain beyond keeping the URL configured. Self-hosting means tracking the openaccountants-mcp package and the packages/ bundles as they regenerate. Contributing means working inside skills/ and waiting for the nightly export, and the README notes that merged source pull requests must be confirmed as ingested before the next platform export. For accountants, the platform commits edits to the repository under a GitHub username set in the profile, so the contribution graph reflects the work. The roster in VERIFIERS.md is generated nightly from the platform, which means a reviewer's standing in the repository depends on the platform record staying current.

Editorial conclusion

Adopt OpenAccountants if you need an AI to cite a tax position that carries a named reviewer and a review date, and you are willing to check the frontmatter before trusting any single Guide. Do not adopt it as a filing engine or as a substitute for a professional opinion on your own facts; the README calls the content general reference, not advice, and warns that Guides may be incomplete, outdated, or wrong. Before wiring it into anything, verify three things: that the jurisdiction you care about is in the accountant-reviewed state rather than a source-cited draft, that reviewed_by resolves to a name on VERIFIERS.md, and that your use of the Guide content fits the OpenAccountants Guide License v1.0 rather than the AGPL-3.0 that covers only the code.

Official sources

  1. License: AGPL-3.0
  2. openaccountants/openaccountants on GitHub
  3. Project website
  4. README
  5. Releases
Community notes

Community notes