Inspiration
I do a lot of vibe coding and my desktop has hundreds of folders and projects. I wanted something that is like Notion or Obsiidian but done automatically for me, doesn't require me to map things and design. Then I thought I could use Codex to build something like it for me but not just that actually make it Codex powered where it automatically scans my desktop, finds projects, creates architectures and summaries and keeps them refreshed - so I can always easily find and remember what the project was about and what it contains.
What it does
OpenCommonplace is a local-first personal operating environment for macOS.
- Workspaces discovers projects only inside folders the user authorizes, then Codex builds cited dossiers and readable architecture maps.
- Notebook gives each day a fresh ruled page while preserving history. 3.Plan holds work that should outlive today.
- Library captures notes, links, files, images, and screenshots, preserves image thumbnails, and organizes everything into visual thematic dossiers with concise summaries. + comes with browser extension where you can right click selection and get it sent to the library
- Newsroom follows RSS and Atom sources, turns incoming stories into digestible themed briefs, and lets the user save only what matters. - e.g. hackernews top news picked and summarised for your via Codex
- Cortex (YOU SECOND BRAIN) connects tasks, projects, captures, decisions, and activity in a navigable knowledge map.
- Command answers questions across an explicit scope and returns citations plus an inspectable receipt.
The core app works locally and does not require an OpenCommonplace account. AI is optional and bounded by the user's evidence-sharing policy. Model output is treated as a proposal: schema, citations, evidence scope, and authority are validated locally before anything is presented.
Why it is different
Most AI productivity products begin with a chat box. OpenCommonplace begins with purposeful tools.
The user does not have to invent the right prompt to get value. They open the Notebook to think about today, Library to remember, Workspaces to understand a project, Newsroom to read, Cortex to connect, or Command to ask a scoped question.
The design also makes a strict distinction between archive and delete. Archive keeps something but removes it from the active view. Permanent delete removes the local record and leaves a suppression marker so sync or rescanning cannot silently resurrect it.
How I built it
The native application is written in Swift 6 and SwiftUI with a modular Swift Package architecture. GRDB and SQLite provide local persistence, FTS5 provides lexical search, Vision provides local OCR, and managed assets are content-addressed and included in portable backups.
The agent path uses a bounded, read-only Codex CLI adapter. Evidence packs are assembled locally, requests are scoped and cancellable, results must match typed schemas, and every accepted result carries inspectable citations and a durable receipt.
The repository includes a Chromium capture extension, macOS Service integration, an optional local Newsroom refresh worker, portable archive import/export, 22 database migrations, full lifecycle tests, and a fail-closed Developer ID signing and Apple notarization pipeline.
For judges who do not want to install a native app, I also built a high-fidelity interactive browser demo with realistic dummy data and clearly labeled simulated AI actions.
How I used Codex and GPT-5.6
I built OpenCommonplace in one continuous Codex Desktop collaboration during Build Week.
Codex helped translate the product manifesto into a working architecture, implement the native surfaces and persistence model, write and run tests, diagnose failures from screenshots and live runtime evidence, improve the product repeatedly, and complete signing, notarization, packaging, and public-demo work.
GPT-5.6 was used throughout that collaboration for the parts that required the most product and engineering judgment: turning intent into system boundaries, tracing cross-layer defects, reviewing security and lifecycle behavior, and improving the evidence-backed agent architecture.
The application itself can also use Codex through the user's existing ChatGPT subscription. Library and Newsroom organization choose a lightweight GPT-5.6 model at low reasoning effort, while deeper cited Project Analyst and Command work use the configured Codex runtime.
The key decisions remained mine: local-first storage, no product account, purposeful surfaces instead of a chat-first shell, explicit evidence sharing, evidence before fluency, archive and delete as separate actions, and a quiet notebook-like visual language. The dated commit history, verification suite, signed release, and notarization evidence make that collaboration inspectable.
Challenges
The hard part was AI behavior trustworthy inside a local desktop product.
That meant handling real filesystem permissions, schema migrations, process boundaries, native browser and Services capture, structured-output failures, stale local databases, lifecycle conflicts, and layouts that looked correct only in one state. Many important improvements came from installing actual builds, reproducing issues against real data, and fixing the root cause rather than hiding the symptom.
What I am proud of
OpenCommonplace is a working, coherent product rather than a static prototype. It has a notarized macOS build, an interactive public demo, durable local storage, real capture paths, evidence-backed project analysis, visual Library and Newsroom organization, a usable Cortex, and a repository-level verification gate.
Most importantly, it demonstrates a different relationship with AI: intelligence can be useful without becoming the interface, and generated answers can be helpful without asking the user for blind trust.
Try it
Interactive demo: https://opencommonplace-demo.sandric.chatgpt.site Source repository: https://gitlab.com/bysandro-group/opencommonplace Signed and notarized macOS preview: https://gitlab.com/bysandro-group/opencommonplace/-/releases/v0.1.0
The browser demo needs no account and uses dummy data. The native preview targets macOS 14 or later.
Built With
- codex
- fts5
- gpt-5.6
- grdb
- sqlite
- swift-6
- swiftui
- vision
