On-device AI by what you need: four categories that don't overlap

Because four different system surfaces are all marketed under the same banner, and they do not overlap. A chat app like Locally AI or Enclave AI runs a model locally and answers questions in a window. It never touches Finder or Mail. A transcription tool like WhisperKit or Parakeet turns voice into text and stops there. A computer-use agent like Fazm reads the macOS accessibility tree of whatever app is open and clicks, types, scrolls inside it; the model can live on-device or in the cloud, the actions stay local. A personal context layer like Fazm's ai_user_profiles or Apple Intelligence's Personal Context indexes your files and identity into a local store that other prompts read from. A reader who needs the third one will not be helped by buying the first one, and most listicles do not draw the line.

A local chat client. Locally AI and Enclave AI are the friendliest macOS apps; both ship a curated model list, both run on Apple Silicon, both have nothing to configure. LM Studio gives you the most control and a model browser tied to Hugging Face. Ollama is the CLI plus a thin GUI; if you already live in a terminal it is the cheapest path. On Device AI ships 170+ Llama, Gemma, Qwen, DeepSeek, Phi, and Mistral checkpoints across iPhone, iPad, Mac, and Vision Pro via GGUF and MLX. Any of those answers 'chat offline' fully. None of them clicks a button in Mail, edits a row in your CRM, or reads the file system to know what kind of work you do; that is the next row down.

A transcription tool, not a chat app. WhisperKit is Apple's on-device port of Whisper, used inside several Mac apps. Parakeet (NVIDIA, MLX port available) is faster than Whisper for English on Apple Silicon and is the right pick when you want low-latency dictation. Superwhisper and EmberType wrap one of those engines into a system-wide dictation experience with a hotkey, a menu bar app, and post-processing. None of them act on the text they produce. Pair one of them with a chat app or with a computer-use agent if you want the transcribed text to actually do something.

A computer-use agent. The two consumer-facing options on macOS today are Fazm and Perplexity Personal Computer. They take different paths. Fazm bundles a Swift MCP server that wraps Apple's AXUIElement accessibility APIs and exposes exactly 9 action tools, each returning a text file where every row is '[Role] "text" x:N y:N w:W h:H visible'. The model grep's that text for the button it wants and passes the bounding box straight back. Perplexity Personal Computer is closer to a chat-driven action layer over your active app. Fazm is consumer software you install from a DMG; you grant Accessibility once and run it. Perplexity Personal Computer launched May 7 2026 and is bundled with their broader Mac client. If you want something you can read the source of and run today with a local model behind it, Fazm is the more direct fit.

A personal context layer. There are two real surfaces here today. Apple Intelligence's Personal Context, rolling out gradually through 2026, ties into Apple's own indices: Mail, Calendar, Messages, Notes, Photos. It is correct for OS-level Siri queries. Fazm's local profile takes a different cut: an AI-generated paragraph profile of you written to a four-table SQLite schema (ai_user_profiles, indexed_files, local_kg_nodes, local_kg_edges) at first launch, by running 5 to 12 SELECT queries against an index of your home folder and installed apps. The profile gets wrapped in <ai_user_profile> tags and pasted into every chat turn before the model sees the message. If you want an agent that already knows the languages you code in by extension count, the apps you have installed, and the projects at the top of ~/, that is the row you need. The data is local. Whether the inference is local depends on the endpoint you point the agent at.

The categories are mutually exclusive in what the software does, not in what you might want to install. A realistic stack for a Mac user who is serious about local AI is a chat app for casual offline questions, a transcription tool for dictation, a computer-use agent for the actual work, and a personal context layer underneath the agent. Fazm covers the last two of those in one app and lets you point the model layer at any Anthropic-Messages-compatible endpoint, including a local MLX bridge. The other rows are still better filled by tools built for them. A computer-use agent is the wrong place to ask a casual question; an offline chat app is the wrong place to expect actions.

Most people who type the phrase mean one of two things and confuse them. Strict on-device means the weights are on disk, inference runs on the Neural Engine or GPU, no network round-trip. That is what a chat app like Locally AI guarantees by default. Loose on-device means the screen, microphone, file system, and personal data stay on the Mac, but the model output may come from the cloud. Fazm in its default config is the second one: the agent loop, the accessibility tree, the SQLite store, the 9 macos-use action tools, and the personal profile are all local. The model is configurable. There is a single field, Custom API Endpoint in Settings > AI Chat, that you can point at any Anthropic-Messages-shaped server, including a local MLX bridge on 127.0.0.1. Sorting by strictness matters because most products in the third row work the loose way, and most products in the first row only work the strict way.

Three open files in github.com/m13v/fazm. Sources/MCPServer/main.swift around line 1482 of the mcp-server-macos-use submodule, which is where the 9 action tools (openAppTool, clickTool, typeTool, pressKeyTool, scrollTool, setValueTool, pressAXTool, setSelectedTool, refreshTool) are aggregated into the tool list the model sees, and line 1505 which documents the row format '[Role] "text" x:N y:N w:W h:H visible'. Desktop/Sources/AppDatabase.swift lines 896 through 928 for the four-table personal-context schema. Desktop/Sources/Chat/ChatPrompts.swift line 101 for the tool-routing rule that hands desktop apps to macos-use, the browser to playwright, and never confuses them.

Pick by the verb. If the verb in your sentence is 'ask', install a chat app. If the verb is 'transcribe', install a dictation tool. If the verb is 'do' or 'finish' or 'send' or 'fill', install a computer-use agent. If you do not know which verb yet, the computer-use agent is the row that subsumes the others later, because once you have an agent that can drive arbitrary apps, you can drive a transcription app or a local chat app from it. The other direction does not work: a chat app cannot grow into an agent without rewriting itself.