Pulse Local Keeps Your Budget Chat and Checks on Your Mac
Pulse Local is a budgeting app with a chat interface that answers from your own figures. Both the ledger, the stored record of your transactions, and the AI models live on your machine. The Mac version is available as of Oct. 13, 2026, with a Windows version planned next. Pulse
No account is required. The app works in any currency and answers in English. CSV import, the standard file export from online banking, works from any bank in any currency. Older Intel Macs are not supported.
First launch pulls the models once. The total is about 2.9 GB. The largest component, at 2.5 GB, checks every money answer. It can be switched off in Settings. With the check on, a money answer takes 12 to 14 seconds.
That wait is the cost of the verification path. Supported spending answers are calculated from your transactions, not generated as free text by the chatbot. Each result is read twice before a figure is shown. When the two readings disagree, Pulse says it is not sure instead of guessing. AI answers are labeled.
Nothing leaves the disk in that loop. The ledger, the models and every question stay on the machine. User questions are not sent to an online AI service.
Network behavior outside that calculation is narrow. Apart from the one-time model download, the only other request checks whether a newer app version is out and downloads it when available. That version check carries nothing about the user and can be switched off in Settings. The stated posture is that nothing else besides the model download and version check is ever sent. Pulse
Optional bank sync uses a proxy pattern, meaning a go-between, rather than direct client to Plaid calls. When syncing, the local app calls Pulse, which asks Plaid, the service that connects apps to banks, for new transactions and returns them in the same response. Transactions pass through the Pulse server's memory and are not written or stored. The server retains only the connection token and the connection record: institution, status and last sync time. Sync brings transactions in and never copies the user's ledger out. Pulse
That sync path is US banks only for now via Plaid. Users who prefer manual input can stay entirely offline with CSV. The connected option is sold as a yearly Sync Pass. Three institutions are included, with additional institutions at $8 each.
The broader context here is the trade that local inference, running the AI on your own device, forces. Keeping retrieval, calculation and a large checker model on device removes routine telemetry, background usage data, and prompt logging from the threat model, the list of risks planners consider, but it pushes storage, memory bandwidth and wait time onto your computer. Twelve to 14 seconds for a checked money answer will feel slow against hosted chatbots. It is fast enough for a ledger query where the failure mode is a wrong number.
In my view, the more interesting design choice is to separate calculation from language. A deterministic read, a fixed rule-based lookup, of local transactions, read twice, with an explicit abstention, a clear statement of uncertainty, when the reads disagree, is a conservative pattern for money. Labeling the AI output keeps that boundary visible. Switching off the 2.5 GB checker trades assurance for speed and disk, and that control belongs in Settings rather than buried in a policy page.
On the sync question, the design follows the same logic on what is kept. A memory-only pass through with no transaction persistence, no saving of transactions, narrows retention, while still leaving a token and connection metadata on the server side. That is a smaller surface, not zero trust, where no server has to be trusted at all. Users and IT reviewers will still need to decide how they weigh a hosted Plaid intermediary against pure CSV import, and US-only Plaid coverage makes that decision straightforward for many non-US households for now.
Looking ahead, what this enables, if the implementation holds, is useful. A no-account budgeting tool that works in any currency, imports from anywhere, and can answer specific spending questions without uploading the ledger lowers the barrier to asking those questions at all. Local-first AI, AI that runs on the device first, has promised that for years without shipping in mundane, high-stakes software like personal finance. This is a concrete attempt to do it where correctness matters more than fluency.
