This repository ships a small MCP server, platform/agents/tools.mjs, that lets an AI agent search the 1,350 calculators and checkers under ventures/, see how to call one, and run its engine without a browser. This run used it the way an agent would, and the first call failed.
The tool was the dog grooming profit calculator. describe_tool showed a DEFAULTS object of ten numbers (card fee, booth rent, commission) and a list of eight exported functions. The obvious first call, compute with those defaults, came back with no exported function "compute"; try: describe. The real entry point is analyze(services, opts), and the DEFAULTS are its options: the services come from pasted text through parseServices, and the module ships a sample of that text in sampleServices(). All of that was in the description. None of it said which call to make.
What changed
describe_tool now returns an example: a call, or a parse step and then a call, that the server has just run on what the tool itself ships and that gave a result. It tries, in order:
- the module's
DEFAULTSas the input of an entry-shaped function (compute,analyze,calculateBid,contributionLimitOutcome, ...); - a shipped sample (
sample*(),example*(), aSAMPLE*constant) as the first argument, withDEFAULTSafter it as options when needed; - sample text through a
parse*function, then the entry function on the parsed rows (or on the array inside a{ rows, warnings }wrapper); - with no
DEFAULTSand no sample, an empty object, but only when the answer has a non-zero number in it; an all-zero result shows the shape and teaches nothing.
A call counts only if it does not throw, returns something, holds no NaN anywhere in the result and does not say ok: false. Infinity is allowed: the grape grower calculator answers "never breaks even at these figures" that way on purpose. The search runs in a child process with an eight-second timeout, so an engine that loops cannot hang the server, and it runs on every describe, so the example cannot go stale.
Two smaller fixes rode along. A wrong function name now lists the real ones (it exports analyze, money, money2, parseServices, ...), which saves the round trip. And every URL the server handed out pointed at the old GitHub Pages address, which since 2026-10-29 serves only redirect stubs; an agent that fetched it got a stub. The server now reads the origin from the site build (https://marisstella.app/), and a test pins the two together.
How many tools get one
node platform/agents/example_coverage.mjs asks every tool. On this run:
| Kind of example | Tools |
|---|---|
shipped sample as args | 336 |
input: {} (DEFAULTS, or the function's own defaults) | 294 |
sample string as text | 107 |
| parse step, then the entry function | 32 |
| none found | 581 |
| with an example | 769 of 1,350 (57.0%) |
The first version, with only steps 1 and 2 and a narrower list of entry names, found 355 (26.3%). An earlier draft of step 4 accepted any answer on an empty object and reached 876; requiring a non-zero number dropped 155 of those, which is the honest number.
The 581 without one are not broken. Most need a real input with no sample in the module (an nginx config, a podcast feed, a set of league teams), or name their entry point in a way no pattern should guess. For those an agent still reads the doc comments, as before.
What next
The cheapest gain is in the tools themselves: a SAMPLE constant or a sample() function next to an engine that takes pasted text makes the tool callable by an agent on its first try, and the page can use the same sample for its "load example" button. example_coverage.mjs --list null names the candidates.
Results from these tools are estimates from the inputs given, not legal, tax, financial or professional advice.