OpenAccountants
Open-source tax guides for AI agents, reviewed by named, licensed accountants — answers your AI can cite on the record.
- Source repo
- openaccountants/openaccountants
- Stars
- ★ 425
- Last updated
- today
- License
- AGPL-3.0
- Primary language
- Python
- FA score
- 75/100 · Good
At a glance
- How it runs
- Works with
- Universal · cross-platformChatGPT · Claude.ai
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A South African freelancer asks Claude what they owe and receives an ITR12 working paper, IRP6 schedule and the reviewer's byline
- Not a fit if
- Users needing formal, signed tax advice for filing decisions
- Teams that cannot send queries through a hosted connector or external platform
- Source review
- 75/100 · Good
What does this agent do, and when should you use it?
OpenAccountants is an open-source tax guide library (AGPL-3.0 for code) spanning 1,881 guides across 230 jurisdictions, of which 53 are accountant-reviewed by 44 named professionals. Every guide sits in one of two greppable states: reviewed by a named licensed accountant (CPA/CA/EA) credited in the frontmatter, or a source-cited draft awaiting professional review. You can consume it three ways: add the hosted MCP connector (https://www.openaccountants.com/api/mcp) to Claude, ChatGPT, Cursor or any MCP client; self-host the MCP server via pip (openaccountants-mcp); or download jurisdiction folders from packages/ and upload them manually. Nightly automation keeps the library honest through a maths check, a source watch, a refresh engine and a nightly exam that catches AI improvisation. Guide content carries a separate OpenAccountants Guide License v1.0 with commercial licensing options.
The repo keeps guide sources in skills/ per jurisdiction; nightly jobs generate packages/ per-country bundles, the machine-readable index., and the llms.txt entry point. Users point Claude/ChatGPT/Cursor at the hosted MCP endpoint or run pip install openaccountants-mcp as a local MCP server; when asked a question, the AI loads relevant guides (e.g. za-income-tax, za-provisional-tax) and produces working papers such as ITR12 and IRP6 schedules with the reviewing accountant's byline (e.g. Werner Britz CA(SA)). A nightly pipeline runs a maths check (tax band gaps, totals that don't total), a source watch (alerts when official pages change), a refresh engine (re-derives stalest guides from official sources on a monthly budget, refusing to publish unlinked figures), and a nightly exam (verifies AIs answered per the guides). Platform edits are committed back under the author's own GitHub identity, and merged PRs flow back into the platform.
- A South African freelancer asks Claude what they owe and receives an ITR12 working paper, IRP6 schedule and the reviewer's byline
- A developer in Cursor queries a 2026 combined sales tax rate via the MCP endpoint instead of letting the model guess from training data
- An accountant authors a guide for their home jurisdiction via openaccountants.com/skills/new, publishing it credited under their name
- A licensed CPA/CA/EA reviews guides in their jurisdiction, putting their name, credential and review date on every AI answer that cites them
- A privacy-sensitive team downloads a jurisdiction folder from packages/ and uploads index. and guide files to their own AI without the hosted connector
How do you install or deploy this agent?
Fastest path: add the hosted connector in your MCP client:
text
https://www.openaccountants.com/api/mcpGuided setup: openaccountants.com/connect. Self-hosted path:
bash
pip install openaccountants-mcpManual path: download your jurisdiction's folder from the repo's packages/ directory and upload the files (including index.) to your AI.
How do you use this agent?
Once connected, just ask, e.g.:
text
What's the combined sales tax rate in Manatee County, Florida for 2026?
The AI loads the matching guides and answers citing the current Guide and the reviewing accountant. Have a qualified professional review outputs before filing, payment, or action.
What are this agent's strengths and limitations?
- Differentiator: each reviewed guide carries a named accountant's name, credential and review date, publicly on the record in VERIFIERS.md
- Broad coverage with flexible delivery: 1,881 guides across 230 jurisdictions, consumable via hosted MCP endpoint, pip self-hosted server, or manual files
- Actively maintained accuracy: nightly maths check, source watch, refresh engine and a nightly exam that catches and fixes AI improvisation at the source
- Only 53 of 1,881 guides are accountant-reviewed; most content is still source-cited draft without professional sign-off
- Explicitly general reference, not advice — outputs must be reviewed by a qualified professional before any filing or payment
- Guide content uses the OpenAccountants Guide License v1.0 (distinct from the AGPL-3.0 code license), with commercial use requiring separate licensing
- The hosted connector sends queries to openaccountants.com; teams that cannot route data through the vendor's cloud must self-host or use manual uploads
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| OpenAccountants This agent | 75 · Good | MCP serverFreemium | ★ 425 | today | Python | ChatGPT · Claude.ai |
| MCP Brasil | 69 · Some gaps | MCP serverFree | ★ 1.8k | 1mo ago | Python | Claude Code · Claude.ai |
| QuantDinger AI Trading OS | 89 · Good | Self-hosted serviceFree + model costs | ★ 12k | 2d ago | Python | Codex · Claude Code · OpenAI API |
| World Monitor | 79 · Good | Web appFree + model costs | ★ 88k | today | TypeScript | — |
How does FollowAgents rate this agent?
Why each dimension lost points
Strong least-privilege and external-effects evidence: the MCP server is read-only over bundled markdown, the container asserts unprivileged uid 65532, HTTP binds loopback by default; source attribution is excellent (reviewed_by frontmatter, VERIFIERS.md roster). Deductions: user confirmation relies only on prompt-level disclaimers, no tool-call confirmation mechanism; no privacy statement for queries flowing to the hosted openaccountants.com endpoint; rollback/degradation paths undocumented.
Failure messaging is a highlight: permission-denied, missing corpus, and slug collisions all have dedicated regression tests requiring fail-closed behaviour with no silent empty catalogue. Self-consistency is good but the hosted stats (1,881 Guides) diverge from the README headline (1,000+). Dependency availability shows only a CI build probe; the runtime dependency list is not in evidence.
Audience and scenarios are explicit (connected-AI users, self-hosting developers, accountant contributors), capability boundaries are clear (general reference not advice, two-state labelling), environment fit is strong (hosted connector, pip, manual files, Docker). Trigger precision is middling: no MCP tool descriptions or routing spec shown.
Information architecture is excellent (clean separation of skills/packages/index./llms.txt, REPO-LAYOUT docs), install notes are thorough, licensing is rigorous (AGPL-3.0 for code, separate content licence). Deductions: no CHANGELOG or release/versioning convention visible; only 53 of 1,881 guides are reviewed, the vast majority are unreviewed drafts; naming stability depends on nightly platform generation outside contributor control.
Output usability is strong: answers carry guide citations and reviewer bylines, examples are concrete. Marginal value is moderate — the core promise is real but must be discounted by review coverage. Cost/benefit is acceptable: both hosted and self-hosted entry are low-friction, but accuracy depends on nightly-sync correctness which static review cannot verify.
Claim traceability is strong: every figure claims a link back to official sources, slug conflicts fail closed, reviewers are checkable. Cross-source corroboration is moderate: the nightly source watch and maths check are described but their code is not in evidence. Fact/inference separation is adequate (reviewed vs draft states), but the draft tier has no independent verification.
- Only ~2.8% (53/1881) of guides are accountant-reviewed; the rest are unreviewed drafts — always check the guide state (reviewed_by field) before an AI cites it.
- The hosted connector sends your queries to openaccountants.com; the repository provides no data privacy or retention statement, so be cautious with sensitive tax data.
- This is a static source review (low confidence) with no executed tests; the nightly sync and maths-check code are not in evidence and their reliability is unverified.
- Most content (packages/, index.) is generated nightly by the platform and manual edits will be overwritten; version history depends on platform behaviour rather than an in-repo CHANGELOG.