Rather than letting a model recall tax rules, this hands it the written procedure for the jurisdiction and marks whether an accountant signed it off or it is only source-cited — which is the distinction that decides how much of the answer you can rely on. Reads need no account, so you can inspect what came back before deciding how far to trust it.
A server over OpenAccountants' library of accountant-written tax Guides. An agent can scope a tax question, load the cited rules for the right jurisdiction, and see for each one whether an accountant reviewed it or it is only source-cited.
- start_help — scope the question first: which jurisdiction, which taxpayer situation, which period
- search_rules — find the Guides that apply, rather than relying on the model's recollection of tax law
- Rules arrive with their sources cited and a quality tier attached, so an accountant-reviewed rule is distinguishable from one that is only source-cited
- Rates and threshold lookups for the jurisdiction and period in question
- Cross-border handling, for questions that involve more than one jurisdiction
- Workflow-read tools covering the procedure an accountant would follow, not just the rule text
- Contribution and verification tools, for adding to the library or signing off a rule — these are the ones that need an account
Nothing to read with: discovery, retrieval, rates, cross-border and workflow-read tools are public. Contribution and verification tools need a free OpenAccountants account over OAuth — calling one anonymously returns an OAuth challenge rather than a silent failure. The hosted endpoint is POST /api/mcp over Streamable HTTP, rate-limited per IP and per signed-in account. The server is also published to PyPI as openaccountants-mcp, so it can be run locally over stdio instead.
One command — pip install openaccountants-mcp