Show HN: Pulse Local, a Mac budgeting app whose AI models run on your machine

PulseOS launched the Mac budgeting app Pulse Local on Hacker News, and its core selling point is that the AI models run on the user's own machine rather than calling a cloud API. The official page positions the product as a "complete budgeting app" rather than a privacy demo: planning, rules, splits, cash flow, forecasting, and other features all run locally.

The cost of this design is written into first launch: the app downloads AI models in one go, totaling approximately 2.9 GB. The largest model, at 2.5 GB, is responsible for reviewing every monetary answer and can be disabled in settings. According to the official description, the app makes only two kinds of network requests—downloading models and checking for new versions—and the latter carries no user information and can be disabled in settings.

Pulse Local can be used without an account, and all features are free. The payment buttons on the page ($0 / $20 / $40) are voluntary support only; the app does not verify which tier the user chose, and every tier leads to the same download. The official explanation is that once the app is installed on the user's machine, it costs them almost nothing to run, so pricing power is left to the user.

The only exception is bank sync. The official description calls this the only feature with ongoing costs, so it is priced separately as a $29-per-year Sync Pass covering three institutions, with each additional institution at $8. The free tier has no real-time sync, and users can only import via CSV.

Bank sync is implemented through Plaid and currently covers only US banks; users elsewhere must import via CSV. The official description of the data flow is as follows: during sync, the app calls Pulse, Pulse then requests new transactions from Plaid and returns them in the same response; transaction data passes through Pulse server memory and is not written or stored, and the server retains only connection tokens and connection records such as institution, status, and last sync time.

This means the promise that "data never leaves the device" does not hold in the sync scenario—transaction data still passes through Pulse and Plaid servers. The official framing uses "not stored" to define the boundary of this path, but this statement is currently only the vendor's own account and lacks independent audit. From the source, Pulse Local's privacy promise depends on users trusting its self-reported server non-storage behavior (this is an inference).

The official page contrasts Pulse Local with most local-first apps, saying that although the latter keep the ledger local, questions are still sent along with related data to a hosted model. Pulse Local's models are files sitting next to the ledger, with no endpoint to send to, so questions also stay on the machine.

This comparison is the vendor's own product positioning, not a third-party review conclusion. It explains why the app downloads a 2.5 GB review model: monetary answers are computed from local transactions and read twice before the number is displayed.

According to the official description, for supported spending answers, the app computes twice before displaying the number; when the two readings disagree, it explicitly says it is uncertain rather than guessing, and AI answers are also labeled. The official description also says monetary calculations were checked month by month against a real ledger for six months, with results accurate to the cent.

These statements all come from the official source, with no reviewable testing details provided. The app's misreading rate for questions and the frequency of "uncertain" triggers are also not quantified. The team says the weights of the small open-source model it tuned have been publicly released on Hugging Face and benchmarked publicly against cloud models, but the source does not show specific scores.

Local mode is aimed globally: the country selector includes 75 countries and supports any currency, but each ledger uses only one currency and never performs exchange-rate conversion. Bank files can be imported in any language, but chat Q&A supports only English.

This combination means non-English users can keep accounts in their local currency and import local bank files, yet cannot ask the app questions in their native language. The official description lists "more bank sync" and "questions the model still gets wrong" as next steps, but gives no release timeline for a Windows version or broader bank sync.

The current version is a Mac (Apple silicon) beta, already signed and notarized by Apple, and it opens like a normal app and self-updates. There is no Windows version yet; the official line is only "tell us, and we'll notify you when it's ready."

Taken together, Pulse Local pushes the privacy boundary to the edge of the device with local models, but the bank sync path still has to cross servers. For users who can accept CSV import, it comes close to "data never leaves the device"; for US users who rely on real-time sync, the privacy promise depends on trusting the vendor's own account.