OpenAccountants: Tax Guides for AI with Named Accountant Review

The Accountant Attribution Problem That OpenAccountants Addresses

AI models can perform tax calculations, but they cannot take responsibility for the answer. A model trained on general tax information produces outputs that lack a named reviewer, a review date, and a traceable source for each figure. When a developer builds a tax assistant on top of a general-purpose AI, the product inherits this gap: the answer may be correct, but there is no accountant whose name is on it.

OpenAccountants addresses this by providing structured tax guides where the most authoritative ones carry the name, credential, and review date of a real, licensed accountant. The guides are stored as files in this repository under skills/, each in a specific jurisdiction and tax type. When an AI assistant loads a guide and cites it, the answer attributes its source guide and names the accountant who reviewed it.

The README illustrates this with a concrete example. A question about freelancer tax obligations in South Africa returns an answer that cites the IRP6 provisional schedule and the ITR12 working paper, and closes with "Reviewed by Werner Britz CA(SA)." Without OpenAccountants, the same question produces an answer from training data with no such attribution.

The project makes the review process publicly auditable. Each guide's frontmatter records the reviewer's name under the reviewed_by key, and the VERIFIERS.md file (generated nightly from the platform) lists all named accountants on the public roster.

Two Guide States and What They Mean in Practice

The README defines exactly two guide states. An accountant-reviewed guide has been evaluated by a named, licensed accountant whose credential appears in the guide's frontmatter and in the public VERIFIERS.md roster. A source-cited draft has been written from primary legislation with each figure cited to its source, but has not yet been professionally reviewed.

This distinction matters for the AI assistant consuming the guides. A source-cited draft may be accurate, but it has not been through the human review that makes the accountant-reviewed state meaningful. Developers building AI products on top of OpenAccountants should check the guide state for each jurisdiction they depend on before deciding how much confidence to place in the output.

The README reports 53 accountant-reviewed guides and 1,870 total guides across 231 jurisdictions as of the last nightly sync. The coverage is broad but uneven: some jurisdictions have multiple reviewed guides while others have only source-cited drafts. The project's contributing documentation explains that accountants can submit or review guides through the platform at openaccountants.com/skills/new, and that their GitHub username links their platform contributions to their GitHub contribution graph.

Connecting an AI Assistant: Hosted Endpoint and Self-Hosted Server

The fastest path to using OpenAccountants is adding the hosted MCP connector. The README shows the URL to add to Claude, ChatGPT, Cursor, Windsurf, or any other MCP client:

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

For self-hosting, the Python MCP server installs from PyPI:

pip install openaccountants-mcp

A Docker image is also available for local development and self-hosting. It runs the MCP server under the Streamable-HTTP transport so that remote MCP clients can connect over HTTP rather than stdio:

docker build -t openaccountants-mcp .
docker run --rm -p 127.0.0.1:8000:8000 -e MCP_HOST=0.0.0.0 openaccountants-mcp

The Dockerfile comment notes that the container defaults to binding on 127.0.0.1 with Docker publishing the port locally, which keeps the service inaccessible from the network by default. Exposing it remotely requires additional configuration documented in mcp/README.md.

For workflows that do not use MCP, the repository provides an llms.txt entry point at the repository root and a per-jurisdiction bundle in packages/, where index.json is the machine-readable inventory. Individual package files can be downloaded and uploaded directly to an AI assistant's context.

How the Repository Stays Current: Four Automated Checks

A tax library that goes stale becomes a liability. The README documents four nightly automated processes that run on this repository.

The math check re-reads every guide and flags sums that do not add up, tax bands with gaps, totals that do not total, and years that disagree with each other. The source watch visits the official pages that supply each guide's figures and raises an alert when a page changes. The refresh engine re-derives the stalest guides from official sources on a fixed monthly budget and refuses to publish any figure without a citation link. The nightly exam tests whether AI assistants citing the guides actually say what the guides say, catching cases where the model improvises rather than quoting.

Changes flow in both directions. Platform edits land in the repository as commits under the editor's own GitHub username. Merged pull requests flow back into what every AI serves, credited to the contributor. A guide with an accountant's byline cannot change without that accountant's one-click approval, according to the README.

Developers contributing to skills/ only need to edit files in that directory. The packages/, index.json, and llms-full.txt files are generated nightly and must not be edited in pull requests.

Limitations: Coverage Gaps and the AGPL License

OpenAccountants covers 231 jurisdictions, but the accountant-reviewed subset of 53 guides is a small fraction of the total. A developer building a tax assistant for a specific jurisdiction may find that jurisdiction has only a source-cited draft, or no guide at all. The README's disclaimer applies to all guides: they may be incomplete, outdated, or wrong for specific facts, and they do not constitute professional advice.

The code components of OpenAccountants (the MCP server in mcp/, scripts/, and tools/) are licensed under AGPL-3.0. This copyleft license requires that any software incorporating those components and offered as a network service must release the source under the same license. Developers building commercial products that include the MCP server should review the AGPL-3.0 terms before shipping. The repository also includes COMMERCIAL-LICENSING.md, indicating that alternative licensing arrangements may exist, though the README does not detail their terms.

Guide content is covered by the OpenAccountants Guide License v1.0, which is separate from AGPL-3.0. The LICENSES/ directory in the repository holds the full text of each license in use.

Alternative Approach: General AI Tax Assistance Without Attribution

The direct alternative to OpenAccountants is relying on an AI model's training data for tax questions. General-purpose models have broad knowledge of common tax rules, and for simple questions in well-covered jurisdictions the answers are often correct. The difference is accountability: training-data answers have no named reviewer, no review date, and no link to the primary legislation that justifies the figures.

For products where the answer's provenance matters, such as a tax filing assistant, a compliance tool, or a service that professionals use for client work, the attribution provided by OpenAccountants is the functional point of difference. For a general information product where approximate answers from training data are acceptable, the additional integration complexity of OpenAccountants may not be worth the trade.

The project is maintained actively: the last push was on 2026-09-27 and the repository receives nightly automated updates. The most recent major release, v2.2.0 from 2026-07-04, added complete metadata, stricter CI, and UK structured rates.

Editorial conclusion

OpenAccountants is the right tool for an AI assistant that needs to cite specific tax figures with a human accountant's name and review date on the record. It is not a substitute for professional tax advice: the README includes a prominent disclaimer that guides may be incomplete, outdated, or wrong for a specific fact pattern, and that qualified professional review is required before filing, payment, or action. The code is AGPL-3.0, which means any server software that incorporates the MCP code must be released under the same license. Check whether the specific jurisdiction and tax type you need has an accountant-reviewed guide or only a source-cited draft before building any workflow that depends on it.

Frequently asked questions

Official sources

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