Pulse Local is a budgeting application for Apple Silicon Macs, offered by its makers as a beta that is signed and notarised by Apple and keeps itself up to date. Its distinguishing claim is architectural rather than cosmetic: the chat that answers questions about your spending is described as running from model files that sit next to your ledger on your own disk, with no endpoint to send a question to.
The team positions this against a pattern it attributes to other budgeting tools. In its account, competing apps either hold the ledger on their servers, or keep the data local while still forwarding the question and the rows that answer it to a hosted model. Pulse Local's stated argument is that a privacy policy promising not to train on your data does not alter the fact that the question itself leaves your machine.
The mechanism is described narrowly. Supported spending answers are calculated from the user's own transactions, and the page says a figure is read twice before it is displayed. When the two readings disagree, the app reports that it is not sure rather than guessing, and AI-generated answers are labelled as such. Labels above a spending answer show which month and category were included, and the user can select one to change it.
The chat is scoped to the ledger rather than open-ended. Documented question types include where money went, comparisons between months, what is left in a budget, and a breakdown "by merchant". The page states the chat reads and answers in English, while bank files can be imported in any language.
Localisation is handled at the ledger level rather than by conversion. The user picks a country once and figures are written in that country's currency format; the page gives 1.234,56 € for Berlin and ¥1,311 for Tokyo as examples. One currency per ledger, and a file in a different currency lands as written, with the import saying so. The page states nothing is converted, ever.
On features, Pulse Local is presented as a complete budgeting app rather than a privacy demonstration: plans and targets, rules, rollover, splits, cash flow and forecast. One claim on the page is that the money maths was checked month by month against six months of a real ledger and matched to the cent. That verification is the vendor's own; neither the ledger nor the reviewer is named in the supplied material.
Pricing is unusually flat. The page says every button ends at the same download, that paid options return to that same download, and that nothing in the app checks which option was chosen. It describes a one-time payment rather than a subscription. The single exception is bank sync, which the team says carries a running cost on its side and is therefore priced separately as a Sync Pass at $29 a year.
The team describes itself as a small group of ML practitioners and product people, and says that once the app is installed it costs them nothing to run — the basis for letting users pay or take it free with every feature included either way. It also says it tunes small open models to answer money questions on the machine, benchmarks them against cloud models in the open, and publishes the weights.
Availability is the sharpest limitation. The build is described as a beta for Mac with Apple Silicon, signed and notarised by Apple, and self-updating. Windows is not available; the page asks interested users to say so and promises to write when it is ready. Bank sync is also described as something being extended to more people, implying it is not universally available today.
For freelancers, designers and developers who keep client and personal finances on one machine, the practical appeal is less the chat than the boundary it draws. A tool that answers spending questions without transmitting the question or the underlying rows removes one category of exposure that comes with cloud-assisted finance apps — relevant if you work under client confidentiality terms or handle receipts tied to named clients.
That reading comes with caveats the source itself supplies. The page concedes the model can misunderstand a question, and that its two-reading check is a mitigation rather than a guarantee. It also lists "the answers the model still gets wrong" among outstanding work. Nothing in the supplied material indicates an independent security review, a third-party privacy audit, or reproducible benchmarks with published numbers.
The open-weights and open-benchmarking claim is the part most worth watching for developers. If the team publishes weights and benchmark results as stated, it gives others something checkable about how a small local model performs on money questions relative to cloud models — a comparison that is otherwise usually asserted rather than shown. The supplied evidence does not include those results, so the claim stands as a stated intention.
The unresolved questions are concrete. There is no stated release date, no version number, and no indication of how large the local model files are or what hardware beyond Apple Silicon is required. The page does not say how bank sync works, which institutions it covers, or what data leaves the machine when sync is enabled — the one feature with a running server-side cost is also the one whose data path is least described here.
What the evidence supports is a downloadable Mac beta with a documented local-inference design, a flat one-time pricing model, and a separately priced sync feature. What it does not support is any claim about accuracy, security or superiority over named competitors, because no such comparison is provided in the source material.
For this audience the sensible reading is a product to evaluate on its own terms: download it, test the chat against your own ledger, and check whether the answers hold up on the categories and months you actually care about. The privacy architecture is the differentiator the team is selling; the budgeting maths and the model's reliability are the parts a user still has to verify for themselves.