For AI & agent builders
Your agent computes the answer
instead of predicting it
Worthune gives an AI agent financial computation to call instead of digits to predict. Over MCP it can list the catalog, read a model\u2019s contract, run any model, and check a numeric claim before stating it. It can also create, update, and project whole households and run their coordination strategies. Every result carries its version, assumptions, and sourced figures.
Language models are good at conversation and poor at multi-year financial arithmetic, and your users cannot tell the difference until it costs them. Give your agent tools that compute instead. It discovers the catalog, reads the contract, runs the model, and cites the version and assumptions in its answer.
Ship it asunified household management, planning software, or calculators & models. One API under all three.
Last reviewed
Twenty tools, two layers
The model layer lists the catalog, reads a model's contract, runs the model, and checks a number before the agent says it out loud. The household layer creates, reads, updates, projects, and decides on a household, reads its decision history, imports a book, and narrates a decision. An agent can carry a whole client, not just one question. Six tools work without a key, and the catalog tools answer keyless for the three sample models. Contracts make tool use self-correcting rather than trial and error.
Answers that show their work
Every run returns its version, its assumptions, any sentinel values it triggered, and the figures it used, with sources. "Per the relocation model v1.0.0, break-even is 16 months" is an answer a user can check.
A verdict that can say no
Each claim comes back with one of three verdicts: it holds, it is violated, or it is out of scope. The last one fires when the claim touches something the model does not compute, so your agent can say "I can't check that" instead of quietly inventing a number. That is the failure mode that actually costs you users.
An answer key for your evals
Your eval stack grades tone, safety, and format. Nothing in it knows whether the break-even is right. Post your assistant's numbers and get a pass or fail per case at the tolerances our own test harness runs.
Know what your model scores
FinPlanBench gives models the published rules and asks for the exact answer. The spread between "within a cent" and "exactly right" is wide, published, and reproducible. It is worth knowing before you ship arithmetic to users.
Something your users can touch
After the math, hand the user a live calculator with their numbers already loaded into the link. The conversation ends in a tool rather than a wall of text.
The agent is about to tell a user three things about a job move. It checks all three first, in one API call, and gets three verdicts:
| breakEvenMonths <= 18 | verified |
| netWealthGain >= 100,000 | violated |
| stateTaxSavings > 0 | out-of-scope |
The first claim held. The second verdict stopped a wrong sentence: the move does come out ahead, just not by the six figures the agent was about to claim. The engine puts the gain at $48,085. The third verdict is the one that matters commercially: the claim is out of scope, so the engine declines to rule rather than guess, because this model doesn’t compute state tax. A guess there would be the most expensive kind of helpful. Every verdict carries the engine’s own value and a record tied to that model version. Sentinel values are part of the contract too: breakEvenMonths = 9999 means never breaks even, so an agent reading the contract won’t report it as 9,999 months.
Watch the film
One MCP endpoint, keyless to try
Streamable HTTP, listed in the official MCP registry as com.worthune/models. Add it to Claude or ChatGPT today, or call it from any MCP client library. Six tools need no key: four for the model catalog, where the three sample models answer keyless, and two that run a whole-household projection on a document the assistant passes in, storing nothing. A paid key opens the rest of the catalog and the tools that operate on stored households, the same records that unified household management keeps current and monitors.
// MCP client config
{
"mcpServers": {
"worthune-models": {
"url": "https://worthune.com/api/mcp/mcp"
}
}
}
// The agent then calls:
// list_models()
// get_model_contract({ model: "relocation" })
// run_model({ model: "relocation", inputs: { ... } })
// verify_claim({ model: "relocation", inputs: { ... },
// claims: [{ path: "breakEvenMonths", op: "<=", value: 18 }] })Questions
Common questions
- How do I stop my AI assistant from hallucinating financial numbers?
- Give it a tool that computes rather than a prompt that asks it to be careful. A language model predicting a compound-interest result token by token has no mechanism for being right. The same model calling a computation endpoint does. Connect the MCP server, and the agent discovers the catalog, reads the contract, and gets an answer it can cite.
- Is there an MCP server that does financial calculations?
- Yes. Worthune's MCP server is at https://worthune.com/api/mcp/mcp over streamable HTTP, and it is listed in the official MCP registry as com.worthune/models. The model tools list the catalog, read a contract, run a model, and check a claim. The household tools create, read, update, project, and decide on a household, and bulk-import a book. Three models answer with no key.
- How can a copilot cite the source of every number it shows?
- Each run returns the version it used, the assumptions the model makes, and every government figure it consumed with that figure's primary source and check date. An agent can surface those directly, which turns “your contribution limit is X” into “your contribution limit is X, from IRS Notice Y, checked on Z.”
- Can an agent check a claim before saying it out loud?
- Yes, with the claim-check tool. Send the inputs and a claim like “break-even is under 18 months” and the engine returns one of three verdicts: the claim holds, the claim is violated, or the claim is out of scope. It also returns its own computed value and a record of the check. Out of scope is the important verdict. It is how the agent finds out the question was outside what the model covers, instead of guessing.
- How do I test whether my finance copilot's math is correct?
- Post your copilot's answers to the grading endpoint and get back a pass or fail per case, at the same tolerances our own test harness holds. Or download the eval datasets and run them in your own test suite. Either way you get an answer key, which most teams shipping financial AI do not have. How the models themselves are tested before release is written up at worthune.com/methodology.
Talk to us