Most AI products assume your data should travel to the model. Local-first tools make the opposite assumption: the model should travel to where your data already sits. For personal work this is not only a privacy preference. Documents, bank transactions and meeting audio are exactly the material people are least comfortable uploading, and they are also the material where a general assistant has the least context to work with.
The eight tools below all put the model next to the data. Some run it on your own machine, some run it inside your browser, and some run it in your own cloud account.
Clean up documents before the model sees them
The unglamorous bottleneck in any document pipeline is conversion. A dense PDF, a spreadsheet or a screenshot handed straight to a model wastes tokens and produces worse answers than clean Markdown would. Sygal converts documents into clean, AI-ready Markdown and structured JSON on your own machine, reading PDF, Word, Excel, PowerPoint, images, scans, screenshots, web pages and YouTube links.
It is a paid one-time license after a fourteen-day trial, and OCR runs through the provider and key you choose. On scans, images and web pages in particular the project claims a large reduction in the tokens your AI needs, which is the argument for doing conversion before prompting rather than after.
A meeting recorder with no bot in the call
Parrot is a free, open-source meeting recorder for Mac with a live AI assistant. It notices when a call starts, suggests answers from your own documents while you are still talking, and remembers every call so you can ask about it afterwards.
The design choices are the interesting part. No bot joins the meeting, because Parrot records what your Mac already plays and hears. Audio, transcription, speaker detection and your documents all stay on your machine, with cloud transcription available only if you switch it on. It is GPL-3.0 with no account and no telemetry, and it requires macOS 14 or later on Apple Silicon.
Small models for professional documents
Nace AI builds small models for professional work rather than general assistants. Its Document Intelligence stack parses, splits, classifies, extracts and grounds thirty-six file types, and every extracted value points back to the page it came from. A decision model called Drex selects the next step in a workflow.
Grounding is the feature worth noticing. An extraction you cannot trace to a source page is a liability in audit and accounting work, which is the vertical the company targets. It sells to enterprises, so there is no public pricing, and the focus is deliberately narrow.
Summarize inside the browser
Apogee is a browser extension that summarizes articles, videos, PDFs, DOCX files and pasted text, inspired by Mozilla's Orbit. It runs on WebGPU or WebAssembly inside the browser, or against your own local Ollama or llama.cpp server if you would rather keep inference out of the extension entirely.
Running in the browser means there is no upload step and no account, and it is available for both Chrome and Firefox. The quality follows whichever model you point it at.
Money that stays on the machine
Pulse Local is a budgeting app for Mac whose AI models run on your machine. It covers budgets and targets, rules, rollover, splits, cash flow and forecasting, and answers questions strictly from your own figures.
The privacy claims here are unusually specific, which is a good sign: the ledger, the models and your questions stay on your computer, bank data is passed through rather than stored, and the app lists exactly which two kinds of network request it makes. The first start downloads the models once, and machines with more memory also get a model that double-checks money answers. Windows is listed as coming.
A trading workspace built around agents
OpenChart is an open-source trading workspace built around charts, alerts and your own AI agents. It supports Pine Script style indicators and a Tea editor for custom studies, and can be downloaded as a desktop build or self-hosted. Agents are part of the design rather than an add-on, which shows up in how the conversation sits next to the chart.
One dashboard your assistant can use
Edi Life OS is a self-hosted dashboard for habits, goals, Kanban boards, personal finance and focus sessions. It ships with an MCP server exposing fifteen tools, so an AI assistant can read and update the same data you do.
That is the pattern worth copying: rather than giving an assistant a separate store of your habits and tasks, expose the one you already maintain. It is MIT licensed, runs with Docker on a PHP 8.1 host, and you maintain the instance yourself.
Serve the models locally
All of the above need somewhere to run. Lemonade is a local server for running models on your own hardware, which is what makes the local-first versions of these tools practical rather than aspirational. If you are assembling a personal stack, a local inference server is the piece that turns several independent apps into one system.
Where to start
If your pain is documents, start with Sygal and then Apogee for reading. If it is meetings, Parrot is the clear pick on a Mac. If it is money, Pulse Local is the most careful about what leaves the machine, and OpenChart covers the trading end. Edi Life OS is the right choice if you want one dashboard that both you and your assistant can edit, and Lemonade or another local server is the substrate all of it runs on.
