OpenAccountants MCP¶
Every assistant answers a tax question from memory, and memory is the worst possible source. Rates change on a budget day, thresholds move with an index, and the answer sounds exactly as confident whether it is current or three years stale. OpenAccountants is the opposite arrangement: a skill library where every figure is attached to the law that produced it, served as twelve MCP tools on a public endpoint with no account and no key.
Server type: remote Streamable HTTP, keyless
Endpoint: https://www.openaccountants.com/api/mcp
Server: openaccountants-mcp 1.4.0 (serverInfo), protocol 2025-03-26
Capabilities: tools, prompts (no resources)
Tools: 12 measured
Registry: io.github.openaccountants/openaccountants 1.0.0, remote declared
Licence: AGPL-3.0 for code, OpenAccountants Guide License for skill content
Rate limit: x-ratelimit-limit 60, resetting each minute (read from the response headers)
Probe: initialize, tools/list, and seven tools/call requests in one session
What the server answered¶
initialize returned 200 with no credential, serverInfo openaccountants-mcp 1.4.0, title OpenAccountants, protocol 2025-03-26, and capabilities tools and prompts with listChanged: false on both. There is no resources capability, so a client should not render a resources panel. The instruction block is where the product states its own terms: every skill sets out the method, names the official source for its figures, and says whether a licensed accountant has reviewed it, across 190+ jurisdictions including US states and Canadian provinces.
tools/list returned twelve:
start entry point: intent + jurisdiction, or it asks for both
search_skills find a skill by keyword or browse
get_skill open one skill by slug
search_rules look up rates, thresholds and rules across jurisdictions
get_deadlines filing and payment deadlines for a country or US state
list_jurisdictions coverage list with skill counts and reviewer names
list_verifiers the named licensed accountants behind the library
get_quote only when the user asks what OpenAccountants charges
validate_vat_number live EU VAT number check
check_audit_exemption audit threshold for a Malta private limited company
compute arithmetic through named, versioned engines
share_with_openaccountants sends feedback or a review request, only on request
Driven: the VAT check returned a real company¶
validate_vat_number on LU26375245 came back 200 with valid: true, country_code: "LU", business name AMAZON EUROPE CORE S.A R.L., address 38, AVENUE JOHN F. KENNEDY L-1855 LUXEMBOURG, checked_at to the millisecond, source: "EU Commission VIES (real-time)", the VIES source URL, and confirm_available: true. The answer does not stop at the yes. It carries the next action in the server's own words: keep the VIES confirmation, number and date, as evidence for zero-rating or the reverse charge, and note that the check confirms the number, not the treatment of the supply.
That boundary is the reason this entry is worth a page rather than a line. A VAT lookup that returned valid: true and stopped would be an invitation for an assistant to then explain the treatment from memory. This one separates the verified fact from the inference and names the escalation path when the user asks for the second one.
The preamble is the product¶
search_rules for the standard VAT rate in MT, IE and DE did not return a rate table. It returned a wrapper followed by 8 passages from 7 skills, and the wrapper is the part to read:
Guidance, not a certified current figure.
Official source for Germany: Bundesministerium der Finanzen.
Authority order: statute / directive > regulations > case law >
official rulings > commentary (persuasive only, never a
citation of record)
Values come from sources you fetch in this session, not from recall.
These are PASSAGES, not a rate table.
Two honest readings of that result. The good one: the server refuses to hand over a number without the citation machinery around it, tells the model to verify against the authority's own site, and names the authority for the jurisdiction it selected. The awkward one: a request phrased as a rate comparison came back with passages about estimated-tax instalments in Malta, not a MT/IE/DE rate table, so the retrieval matched the jurisdictions and not the intent. An operator asking "what is the VAT rate in Ireland" should expect the rate to arrive from get_skill on the VAT skill, with search_rules as the supporting citation rather than the answer.
Coverage, and the field that reads zero¶
list_jurisdictions returned a per-country table. The top of it, read as counts of skills:
GLOBAL 83 US 73 AU 60 ES 53 CA 47 GB 34 MT 34 DE 30
NL 28 FR 23 DO 22 ZA 22 IT 21 IN 20 GI 19 MA 19 PT 19
Named reviewers appear beside some of them: Edgar Lautsyus for Canada, James Power for Great Britain, Michael Cutajar, CPA (Malta) for Malta, Miguel Lantigua for the Dominican Republic, Werner Britz for South Africa and Mayur Deokar for India. And then the field that does not agree with any of it: accountant_verified_skills reads 0 on every single jurisdiction, including the six with a named lead verifier. The README's own banner claims 33 accountant-reviewed skills. Whatever the intended meaning of the field, a per-country approved count that is zero everywhere is not an answer to "who reviewed this", and the guide records it as a defect of the listing rather than as an absence of reviewers, because lead_verifier and lead_verifier_credential carry real names.
start with no arguments asked for an intent and a jurisdiction and returned both lists: 16 intents from income tax and VAT/GST through payroll, formation, crypto, transfer pricing, capital gains, tax residency, non-dom and remittance, special tax regimes and exit tax, and roughly 180 jurisdiction codes, including sub-national ones (CA-AB, EG-AE, EG-SA, EU-27, and US states).
Deadlines: exact where it is exact¶
get_deadlines for DE with the default 120-day horizon returned upcoming_deadlines: [] and a single recurring_rhythms entry, de-ustva-monthly for the monthly Umsatzsteuer-Voranmeldung, with a link to the public tax calendar. That is the right shape for a jurisdiction whose monthly filing has no fixed calendar date inside the window, and it is honest output rather than a filled table. Asking again with a longer horizon is the documented move; a client that renders an empty list as "no obligations" would misread it, so the response carries its own next_action telling the model to widen the horizon or report the rhythm instead.
The calculator is public code and a broken door¶
compute is the most interesting tool and the one that does not work over MCP today. It exposes 15 named engines with explicit versions (de_est_tariff_2026, us_federal_income_tax_2026, us_se_tax, uk_paye_take_home_2026_27, ca_income_tax_2026, au_resident_tax_2026_27, mt_income_tax_2026, nl_vpb_2026 and others), and each needs_input reply carries an inspect block with the engine's code_url, tests_url, figures_url and the exact command to rerun its tests, so the arithmetic is auditable rather than asserted.
The instruction on that tool is the design decision worth copying: never compute the figure yourself instead. The model is told to ask for the missing fact and to stop, not to estimate.
The defect is reproducible. Called as compute(engine="de_est_tariff_2026", taxable_income_zve=60000, assessment="single", tax_year="2026") it returned needs_input with what: "missing input 'tax_year'". Called as compute(engine="us_se_tax", schedule_c_net_profit=60000, tax_year="2026") it returned the same message naming the same field. Two engines, the field supplied both times as the string the schema's own enum demands, the same refusal. So the calculation path is not reachable through the hosted MCP in 1.4.0; the engines themselves are in calculators/engines/ with tests and can be run with node --test. Recorded as a defect with the reproducer, not as a missing feature.
One smaller inconsistency from the same session: the slug de-ustva-monthly returned by get_deadlines is not a skill slug, and get_skill on it answered No published skill 'de-ustva-monthly'. The two tools use different slug namespaces, so a client chaining them will fail until it routes through search_skills first.
The licence decides who can use it¶
Code is AGPL-3.0-only. Skill content is under its own OpenAccountants Guide License, with commercial options documented in the repository. The README states its status plainly: general reference, not advice, and skills may be incomplete, outdated or wrong for the facts of a particular filing. For an agent workflow the honest use is a cited first pass that a qualified person signs off, which is exactly what the server's own instruction block tells the model to say.
What it changes for an operator¶
The gap it fills is narrow and real: an assistant that can say "the standard rate for this jurisdiction is X, under article Y, as published at Z, checked on this date" instead of a remembered figure. Beside the existing accounting entries here, the division of labour is clean. mighty-accounting-mcp works a company's own books, pagofix-mcp drafts the late-payment interest a client owes, honest-vies-mcp validates a VAT number against VIES with its own evidence standard, and casus-mcp holds a French tax and legal corpus. OpenAccountants is the cross-border rulebook the others do not carry: 249 jurisdictions, every figure required to name its law, and a reviewer named where one exists.
The limits, stated the way the server states them. Keyless means rate-limited (60 requests a minute read from the headers), the library is guidance and not advice, every figure should be confirmed against the authority named in the passage, accountant_verified_skills reads zero per jurisdiction, the compute path is broken in 1.4.0, and two tools disagree about what a slug is. What remains is still the strongest answer in this catalogue to a tax question that must be cited rather than recalled.
Integration with CorpusIQ¶
CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.
This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.