A wrapped Wolfram Alpha endpoint for agents, in build and not yet sellable: a factual answer across roughly forty curated domains, exact Wolfram Language evaluation, a chart returned as a PNG — and a citation with every result.
What it will do
The service sits in the studio’s parts bin as external/wolfram_alpha v0.1.0. It is in build. There is no endpoint to call, no price quoted and no date set; what follows describes what has been specified, not what has been exercised. That distinction is this review’s method, and it is stated here rather than at the end.
Three capabilities are specified, and they are separate in kind.
- A factual answer across roughly forty curated domains. Domains the studio’s own endpoints do not cover: the fleet answers on models, markets and on-chain state, not here. The description does not name the domains, and this review does not guess at them.
- Exact evaluation of Wolfram Language. An expression is evaluated to a value rather than estimated at one. The caller supplies the computation; the engine performs it.
- A chart, returned as a PNG. An image artefact: a file to place, not code to run.
What holds the three together is the return. Every result carries its interpretation (how the engine read the question), the link to that specific result page, and the attribution Wolfram’s terms require — the credit the upstream licence demands. None of the three is optional, and none is assembled by the caller afterwards.
The service is a wrapper. x402.press would wrap an upstream Wolfram API rather than reimplement it, which fixes what can be offered: coverage, exactness and terms are Wolfram’s; the payment plumbing and the shape of the call are the wrapper’s.
Table 1. The three capabilities as the parts-bin description states them. Source: external/wolfram_alpha v0.1.0, 30 October 2026. No request schema or field name is published.
| Capability | What the caller brings | What comes back |
|---|---|---|
| Factual answer | A question in words | An answer, its interpretation, the link to that result page, attribution |
| Wolfram Language evaluation | An expression to evaluate | An exact value, its interpretation, the link, attribution |
| Chart | A question or expression where a chart is wanted | A PNG artefact, with the interpretation, link and attribution above |
Why an agent needs it
A language model answers in the register of a fact. The prose is fluent, the number is specific, the sentence shaped exactly as a correct sentence would be. That is what makes a wrong figure expensive: it does not look wrong, so the only way to find it is to check it.
For a draft a person will rewrite anyway, that cost is small. For a document that leaves the building — a market note, a client report, a projected slide — someone must verify every number before it goes. Verification is the expensive part. Drafting is the cheap part, and it is the part that has been automated.
So the saving is not automatic. An agent that produces a document a person must check line by line has moved the labour, not removed it; if every figure has to be traced by hand, the labour that remains is the labour that mattered. What closes that gap is not more confidence but a route back: a way for a reader to reach the page the figure came from without re-deriving the argument.
Three things make a figure usable.
Interpretation. A model that misreads the question still returns a fluent answer, and the misreading is invisible in it. A computed result exposes the reading it acted on, so the caller can disagree before relying on the output.
Provenance. A link to the specific result page, not to the engine’s homepage. A citation that requires the reader to re-ask the question is a reference, not a source.
Attribution. The credit Wolfram’s terms require binds anyone republishing the result, and doubles as an audit trail: it names the engine and the terms under which the figure may travel.
This is a narrower claim than it may sound: not that the service answers better than a model in general, nor that it replaces one. For most of what an agent writes — framing, explanation, summary, argument — a model is the right instrument and no citation is wanted. The claim is confined to questions where an exact value and a checkable source matter more than fluent prose: there the engine computes, and the answer arrives with the page it came from.
Agents work at volume: ten claims an hour is a normal rate, and no person is in the loop for each. The only check that scales is one that resolves to a page — a page a person can open in a second, or another agent can open with no person at all.
How an agent would use it
The call is a paid HTTP request. What follows describes the shape of the call and the shape of the return; no request schema, path or parameter name is published, and none is invented here.
Checking arithmetic before it reaches a document
The agent has derived a figure — a rate, a ratio, a total — and is about to write it into a paragraph. Before it does, it sends the computation itself for evaluation and compares the returned value with its own.
If the two agree, the paragraph keeps the number and gains a citation. If they differ, the draft stalls, which is the purpose of the check: a disagreement is unambiguous, not a matter of degree. The cost is one call per derived figure, confined to figures that appear as distinct claims rather than to every intermediate step.
Putting a cited figure in a report
The agent reaches a fact it does not hold: a rate, a conversion, a date, a ranking. It can write around the gap, guess, or ask for a computation. The third produces a sentence with a source attached: the answer, the interpretation showing how the question was read, and the link, which becomes a footnote.
The interpretation is what a careful agent reads first. If the engine read “per year” as “per month”, the answer is precise and irrelevant, and the interpretation is the only place that shows.
Producing a chart for a deck
A deck needs a visual. Asking a model for one yields code that may not run or may plot the wrong series, with no way for the reader to tell which. Asking for the chart returns a PNG: a file to place, with the same interpretation and link that accompany an answer.
The discipline carries over. A chart is an argument about numbers, and a chart without its working is a decoration; the interpretation is what lets the slide’s author say what it shows rather than what it looks like.
In each pattern the agent does two things: it decides the claim is one where an exact value and a source are wanted, and it checks the interpretation before using the result. The rest — payment, request, parsing — is ordinary HTTP work.
A worked example
Illustrative, not a measurement. No price exists, so the arithmetic here is about when to call, not what a call costs.
An agent drafts a six-hundred-word market note. It makes six factual claims:
- the size of a market in a named year;
- its growth rate between two years, which the agent derived from two figures;
- the date a rule took effect;
- the ranking of a country on a named measure;
- a projection for a later year;
- a comparison between two regions.
Its corpus covers claims 3 and 6 from material it was given. Of the rest, claim 5 is a forecast and claim 1 a market estimate belonging to whichever research house produced it; neither is computable, and routing either to computation would be a category error. Claims 2 and 4 are computable.
The decision rule is three questions, asked of each claim in order:
- Is it computable? A number, a date, a conversion, an evaluation, a property of a defined thing. A judgement, a forecast or an opinion is not: the sentence should be rewritten so it does not pretend to be a fact.
- Is it inside the curated domains? If coverage does not reach it, the claim goes back to the corpus or is marked unsourced and escalated. Absence is a routing answer, not a failure.
- Will a reader act on it or quote it? A figure that will sit in a table, be repeated in a summary or be relied on in a decision earns a citation. Decoration does not.
Two claims pass all three and are routed to computation: the derived growth rate, evaluated rather than estimated, and the ranking, answered with its page. The document changes in two places. Two numbers that were assertions are now citations: a reader can open the page behind the growth rate and the page behind the ranking.
What the review step gains is specific. Without the two calls, a reviewer faces two figures with no provenance and two choices: reconstruct them, or let them pass. With the calls, the reviewer opens two links and reads two interpretations. The note ships either way. The difference is whether its figures can be checked in a minute or not at all — and whether the reviewer is verifying a document or auditing its author.
The cost side is trivial: the price is per call and fixed once quoted. The question an operator must answer is not whether two calls are affordable but which claims deserve one — and it has the same answer whether the endpoint exists or not.
Where it stops
Curated domains only. Roughly forty of them, unnamed in the description. Anything outside that set is not answered, and an agent must hold a fallback: a service that declines differs from a model that improvises — silence is safer, but still silence in the middle of a draft. A widely read agent will find the edges quickly.
It computes what is computable. Not judgement, not strategy, not whether a course of action is sound. The awkward case is the question that looks computable and is not: a precise answer to a slightly different question, returned with full confidence because the computation was correct. The interpretation is the only defence, which is why it is part of the return and not an extra.
Attribution is an obligation. The attribution Wolfram’s terms require comes back with every result, and passing it into the published document is not a courtesy. A figure used without it is used outside the terms it arrived under; the same applies to the link, which is part of the deliverable and not decoration.
A wrapper inherits the upstream. Coverage, exactness, quotas, rate limits, terms and availability belong to Wolfram. The wrapper can add payment, packaging and a uniform return; it cannot add a domain, lift a ceiling or promise continuity of an interface it does not control. A buyer who needs a contractual assurance on any of those is buying at the wrong layer.
Nothing is live. There is no endpoint, no price and no date. No latency, uptime, accuracy figure, quota or domain list is published, and this review supplies none. The three capabilities are as described, not as exercised, and no part of this article is a measurement.
When not to use it
The question is opinion. Which market to enter, whether a strategy is sound, what a client ought to do with a position. No computation returns that, and a precise-looking answer to an opinion question is worse than an argument, because it borrows the authority of arithmetic for a judgement that has none.
The answer must come from the client’s own documents. A contract term, an internal figure, a policy in force at a particular firm, a number from last quarter’s ledger. The engine knows the world; it does not know the file, and a call that returns the world’s version of a client’s number returns the wrong answer confidently.
A model’s answer is good enough and no citation is required. Internal drafts, brainstorming, tone, restructuring, summarising text already in hand. A paid call and a footnote on a paragraph nobody will quote are friction with no return.
What remains, once those three are set aside, is the case the service is for: a figure that will travel — quoted, filed, projected, relied on — where being right is necessary and being checkable is the point.
How it will start
The intended shape, once live, is one paid call: the caller sends a question or an expression; the response carries the answer or the exact value, the interpretation, the link to that specific result page, and the attribution the terms require — with a PNG artefact where a chart is asked for. Payment is the only thing between an agent and the result; nothing is assembled afterwards, because the citation travels with it.
That is the design as it stands in the parts bin, not a running endpoint. Nothing responds to a request today; no price has been quoted and no date set. When that changes, this review will be rewritten against a live call and these claims retested against it.
One piece of preparation does not depend on the interface: the routing rule — which claims an agent computes, which it sources from its corpus, and which it declines to state. It can be written now and will survive whatever the request looks like.
Method note
The description in external/wolfram_alpha v0.1.0 was captured from the parts bin on 30 October 2026 and is the sole source for the capabilities above: the answer across roughly forty curated domains, Wolfram Language evaluation, and the chart as a PNG, with interpretation, link and attribution on every result. On that date no endpoint existed, no price had been quoted and no launch date set. The service is in build and not yet sellable.
No request schema, parameter name, field name, domain list, latency, quota, uptime or accuracy figure was available, and none is inferred. The upstream terms were not examined; the attribution obligation is reported as the description states it. The worked example is constructed, not measured, and contains no figure from any service.
The claims made are deliberately narrow: that the service computes and cites, and that a citation is what makes a computed figure usable. No claim is made that it replaces a model, or that it is more accurate in the abstract.
Filed under Service Reviews · Tags: x402, wolfram-alpha, deterministic-compute, citations
Related reviews: DeepSeek review · Qwen review · HyperFusion review · SpiralStake review · ElevenLabs review · Mage review
Background: what x402 is, and how a payment works
All seven reviews: the services reviewed
