Contha MCP¶
A sole proprietor keeps two sets of money in one life. The personal card buys the laptop. The business account pays the contractor. The tax return wants the second sorted by line, the household wants the first sorted by category, and most tools pick one side.
Contha keeps both sides in one set of books and exposes the whole thing over MCP: 22 read tools that answer what is left to spend, what is owed in tax, where the money went, and which charges were filed on the wrong side. The write tools exist on the production endpoint and are deliberately absent from the demo.
Server type: remote Streamable HTTP (hosted)
Demo: https://demo.contha.com/api/mcp no credential, sample books
Production: https://app.contha.com/api/mcp OAuth 2.1
Server: contha-demo, version 0.4.0, protocol 2025-06-18
Auth: demo none; production OAuth with PKCE S256
scopes read, write; dynamic client registration open
Tools: 22 exposed on the demo, all read-only
22 read and 6 write on the production service, per the vendor
Registry: no official registry entry this pass
Queue: mcp.so #7627, filed 2026-11-18T02:34:39Z
License: MIT, github.com/noahowsh/contha-claude
The evidence tier¶
Keyless and driven, which is rare on this class. initialize answered 200 on the demo host with no header of any kind and returned serverInfo contha-demo 0.4.0 on protocol 2025-06-18. tools/list returned 22 schemas. Fourteen of the twenty-two were called and answered with live data, and three failure paths were called on purpose.
The production host is the opposite shape and states it plainly:
POST https://app.contha.com/api/mcp 401
www-authenticate: Bearer realm="contha", error="invalid_token",
resource_metadata="https://app.contha.com/.well-known/oauth-protected-resource"
Both documents resolve on the same host, which is the cleanest OAuth deployment seen on this surface for weeks: the protected-resource document names scopes read and write with bearer_methods_supported: ["header"], and the authorization-server document carries authorization_endpoint, token_endpoint, registration_endpoint, response_types_supported: ["code"], code_challenge_methods_supported: ["S256"] and token_endpoint_auth_methods_supported: ["none"], which is a public client with dynamic registration and no secret to paste.
No customer books were read, because no account was opened. Everything below is either the demo's own sample books or a surface any reader can re-measure in two requests.
Two books, and the audit that proves it¶
get_overview answers the operator question first: $40,380.70 free to use, against $60,875.91 of net worth split into $45,509.84 cash, $17,244.34 invested and −$1,878.27 owed on cards. The tool prints its own working, and the arithmetic closes to the cent:
cash 45,509.84 − cards 1,878.27 − goals set aside 1,595.50
− bills due in 30 days 296.38 − tax due on the return 1,358.99
= 40,380.70
This month splits the same way: $6,500.00 in business income against $366.94 of business spend, and $508.79 of personal spend, for $875.73 spent on both sides together.
The tool that earns the integration is audit_misfiled_business. It returns charges from known business vendors that are either uncategorised or filed on the personal side, largest first, and on the demo books it returns five findings totalling $115.98: one OfficeMax charge at $55.98 and four Figma charges at $15.00 each, every one with a null category and the matched vendor named. Fifty-five ninety-eight plus four times fifteen is one hundred and fifteen ninety-eight, and that is the whole point of the tool: a business expense that never reaches the Schedule C is a deduction an operator pays tax on.
The tax worksheet, and the one number that does not close¶
get_schedule_c returns a 2026 Schedule C built from the transactions: gross receipts $71,500.00, total expenses $11,213.65, tentative profit $60,286.35. Four lines carry the expenses and they sum exactly:
line 9 car and truck 91.20 2 transactions
line 11 contract labor 7,258.99 10 transactions
line 18 office expense 166.32 11 transactions
line 27a other expenses 3,697.14 55 transactions
---------
11,213.65
get_tax_estimate then walks the same income through self-employment tax and reproduces the standard arithmetic: net profit $60,286.35, self-employment taxable earnings $55,674.44, Social Security at 12.4% giving $6,903.63, Medicare at 2.9% giving $1,614.56, combined self-employment tax $8,518.19, half of it deducted at $4,259.10, adjusted gross income $56,027.25, the qualified business income deduction $7,985.45, federal taxable income $31,941.80, federal tax $3,585.02 and New York State tax $2,428.47 on the 2026 IT-2105-I rates with the recapture worksheet.
The ambiguity is worth an operator's attention rather than a gloss. Three figures sit in the same response and they disagree:
headline $1,358.99 due with the return on Apr 15 (and repeated in get_overview)
balanceDue −$1,668.32 tax on income so far, less $16,200 paid
quarterly toPay 0.00 on all four installments, $1,511,232 required,
$1,620,000 paid, sum of installments $1,511,232 against
requiredAnnual $1,511,233
The headline is internally consistent: get_overview and get_tax_estimate both print $1,358.99. It is not reproducible from the fields published beside it, and a fourth reading (projected full-year $16,791.48 less $16,200.00 already paid, which is $591.48) fits neither. Recorded as measured, not reconciled. The tool says its own limits in every tax response: these are estimates from the transactions and the situation on record, not tax advice, and the exact figure belongs with an accountant before anything is paid or filed.
Units, defaults and the budget header¶
Three conventions are worth knowing before wiring this into an agent, because each one changes an answer.
Money arrives twice: bare integers in cents, and a readable string. 12345 is $123.45, and most objects carry both, so an agent that quotes the integer reports a hundredfold error. The tools that matter state the rule in a units field rather than leaving it to the reader.
get_budget pools carried-in money into its header. Planned $1,695.00, spent $875.73, and leftToSpend $1,869.27, which is not the subtraction: the Equipment line carries $1,050.00 in from earlier months, and 875.73 + 1,869.27 = 2,745.00 = 1,695.00 + 1,050.00. A reader who checks the header against planned − spent finds a figure $1,050.00 too high and concludes the tool is wrong. It is right, and the carriedIn field on the individual line is where the extra money is declared.
Defaults are dated rather than lifetime. get_cashflow with no arguments returned 2026-08-21 to 2026-11-18, income $19,500.00, spend $4,200.33, net $15,299.67, with the first and last months marked partial and their covered ranges printed. get_spending_trends with no arguments returned January 1 to today across eleven months. A question asked without a range gets ninety days of cash and a year of categories, which is two different windows from one conversation.
get_year_end_package takes a kind and returns CSV text with the filename on line one and the line descriptions in column two, so a Schedule C export reads 1,Gross receipts,,71500.00,71500.00. The enum is exact: transactions, schedule-c, mileage, k1, 1099, pnl.
Can an agent see both sides of the books?¶
Yes, and that is the design. Every tool that reports money carries a side or a scope: get_transactions filters on the side of the books, get_budget returns both sides with a side per line, get_cashflow and get_spending_trends take combined as the default scope. The Schedule C tools read the business side only, and get_profit_and_loss states its own basis: cash basis from the business books, which differs from the Schedule C profit because the return adds mileage and leaves out charges never filed to it.
What does it cost, and what does the free plan cover?¶
The vendor publishes a free plan with Personal at $5 a month and Business at $10 a month. The Business plan is the one that carries the personal and business split, the profit and loss, and the Schedule C worksheet. The demo endpoint used in this guide is open with no login and reads sample books only, so every number above is reproducible without an account and without a payment method.
What can the agent actually change?¶
On the production service, six write tools file transactions, add cash transactions, and maintain rules, budgets, mark bills and goals. The vendor states the boundary directly: it can never move money. The demo exposes none of them, and a call to a production-only name returns Unknown tool: add_cash_transaction rather than a permission error, which is the cleanest way to see the two surfaces are genuinely different builds.
Connecting¶
The demo, for evaluation with no account:
{"mcpServers": {"contha-demo": {"type": "http", "url": "https://demo.contha.com/api/mcp"}}}
The production endpoint, for real books:
{"mcpServers": {"contha": {"type": "http", "url": "https://app.contha.com/api/mcp"}}}
In Claude, Settings then Connectors then Add custom connector, paste the URL, and press Allow on the page Contha opens. The consent screen offers read and write separately, and the vendor says write needs explicit consent at that page. Dynamic client registration is open, so there is no client id and no secret to request first.
Two cautions for an operator. The demo answers with a full set of books that are not yours: state that plainly in any answer an agent builds from it. And the tax tools are estimates with a fixed prompt to say so, which is the right behaviour and still a statement that needs a human signature before money moves.
Sources¶
- Demo endpoint: https://demo.contha.com/api/mcp
- Production endpoint: https://app.contha.com/api/mcp
- Protected resource metadata: https://app.contha.com/.well-known/oauth-protected-resource
- Authorization server metadata: https://app.contha.com/.well-known/oauth-authorization-server
- Vendor AI documentation: https://contha.com/ai
- Repository: https://github.com/noahowsh/contha-claude
- Submission: the mcp.so queue as #7627 on 2026-11-18
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.