zihenghe04/AIBro

AI Bro is a local-first Mac workspace that treats a conversation not as a disposable exchange but as the beginning of a persistent work record. The core design premise is that most AI tools produce outputs you immediately lose track of — answers that drift away from their source documents, edits you cannot verify against the original, plans that have no home once the chat window closes. This app tries to close that loop.

The mechanism is a layered workspace: files, notes, projects, a research wiki, and a calendar all share the same context. You drag in a paper or a slide deck, work through it with an AI, and the result — the diff, the annotation, the follow-up question — stays attached to the original file inside a project. A side-by-side reader keeps the source document visible while you work, so you are checking evidence rather than trusting a summary. The diff review step is deliberate: AI-generated edits are shown as discrete, reviewable changes before anything is saved.

For research specifically, the wiki layer is the most interesting part. It is designed to accumulate not just conclusions but failed experiments, open questions, and review feedback — the material that normally lives in someone's head or scattered across a folder of PDFs. Concepts can be linked bidirectionally, and sources can be traced back through those links. That is a meaningful structural choice: it treats research as an ongoing record rather than a series of one-off queries.

The agent capabilities are practical rather than elaborate. Local BM25 keyword indexing handles retrieval by default, with optional vector search and reranking. File tools, web search, terminal commands, and a research sub-agent are all available, with an approval step for consequential operations. Skills let you save and reuse common workflows. None of this is novel in isolation, but the combination of agent actions with a persistent project context — where the project holds a plan, a daily log, and long-term memory — gives the agent something to work against beyond the immediate conversation.

The iOS companion is unsigned IPA requiring your own developer certificate to install, which is a real friction point for anyone outside the developer community. The University of Chinese Academy of Sciences course assistant feature — attendance check-in, schedule import, dynamic QR codes — is clearly built for a specific student audience, and it shows in the overall product framing. The README is in Chinese first, English second.

Data stays local. You choose which model API to connect, which directories to authorize, and whether to run a self-hosted sync service for cross-device continuity. When you use a remote model, only what is needed for the request goes to your configured provider. The licensing is AGPL-3.0.

The build stack is a hybrid: SwiftUI and AppKit for native Mac UI, with npm and Python dependencies suggesting a mixed native-and-web-technology architecture. The macOS 26 Liquid Glass styling is a cosmetic detail, but it signals the project is tracking platform conventions closely.