Skip to content
architpaiPublic

About

A native macOS notch companion for supervising Codex work.

Topics

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Conn

Conn — a notch-native supervision surface for AI harnesses

Visit the Conn website · Download the latest release

Conn is a native macOS notch companion for supervising AI harness Sessions while you work in other apps. It shows connected activity, surfaces supported permission and question requests, and exposes only the controls that each Integration can safely support without taking ownership of the harness process.

Important

Conn 0.2.1 remains an alpha preview for Apple Silicon Macs running macOS 15 or later. Codex 0.144.5 and 0.144.6 and Pi 0.83.0 are the only qualified harness versions. Both Integrations are off by default and must be enabled explicitly in Settings. Shared Desktop Mode remains experimental.

Conn is an independent open-source project and is not an official OpenAI product.

What Conn does

  • Keeps the Sessions in the current active and 24-hour view visible in a compact top-center surface, with status counts scoped to exactly that visible set.
  • Expands into a focused workspace with harness attribution, Session switching, chronological activity, grouped Runs, full completion summaries, and elapsed work time.
  • Supports capability-gated approval, question, follow-up, steer, interrupt, model, and reasoning actions through Codex App Server.
  • Observes independently launched Pi TUIs through Pi's standard global extension and supports follow-up, steer, interrupt, model visibility, and idle model switching. Pi has no standard approval or question controls.
  • Starts a Codex Session as a local draft in the default Workspace, with one compact model-and-reasoning control; Codex creates the real Session only when the first message is sent. External Pi Sessions must be started in Pi.
  • Lets users enable Codex, Pi, both, or neither without changing the lifecycle of work already owned by either harness.
  • Rehydrates bounded state after reconnecting without taking ownership of Codex or Pi Sessions and their lifecycle.
  • Offers an optional Labs flow for qualifying Codex Desktop and Conn as clients of the same managed daemon.

Requirements

  • Apple Silicon Mac. Intel builds have not been validated yet.
  • macOS 15.0 or later.
  • At least one qualified harness:
    • Codex installed, authenticated, and exposing CLI/App Server version 0.144.5 or 0.144.6.
    • Pi 0.83.0 installed with the Node runtime that provides the pi command.
  • Swift 6 and the Xcode Command Line Tools only when building from source.

Install

Download the latest macOS archive from GitHub Releases, verify the included checksum, open the DMG, and drag Conn into Applications.

The current alpha build is ad-hoc signed. macOS will therefore require an explicit Control-click Open on first launch, and Launch Conn at login cannot be registered by this artifact. Developer ID signing and notarization are planned before Conn leaves alpha. See INSTALL.md for the exact binary and source installation paths, Gatekeeper steps, supported versions, verification commands, and uninstall instructions.

Five-minute smoke test

  1. Confirm an authenticated Codex 0.144.5 or 0.144.6, Pi 0.83.0, or both are installed.
  2. Install and open Conn using INSTALL.md.
  3. Open Settings and explicitly enable the Integration you want to test. Pi setup installs Conn's extension after confirmation; already-open TUIs need one /reload.
  4. For Codex, leave Shared Desktop Mode off and start or resume a harmless Session through the managed daemon. For Pi, start a harmless TUI Session.
  5. Confirm the Session appears with the correct harness mark, model, status, and transcript.
  6. Send a benign follow-up or steer while active, then verify interrupt on a disposable Run.
  7. For Codex, verify draft-first New Session. For an idle Pi Session, verify an available model can be selected before the next follow-up.

Conn fails closed when a harness version is unsupported. Approval and question controls only appear when Codex emits a matching request and grants Conn authority to answer it; Conn does not invent equivalents for customized Pi tools.

Architecture

Conn is a Swift 6 menu-bar/accessory application. Its normal integration path has two adapter-owned paths:

Conn.app -> codex app-server proxy --sock -> Codex-managed App Server daemon
Conn.app <- bounded local socket <- Conn Pi extension <- external Pi TUI

The Codex proxy is disposable; Codex owns the daemon, threads, and turns. The Pi extension reports bounded structured state and accepts acknowledged commands; Pi owns its TUI, Session, and runtime. Quitting or disabling Conn does not stop either harness.

The provider-neutral implementation is split into:

  • ConnDomain for Harness, Integration, Session, Run, Activity, Attention, and action semantics.
  • ConnAppCore for Integration aggregation, persistence, presentation, and policy.
  • ConnCodexAdapter for the version-gated Codex App Server implementation.
  • ConnPiAdapter for the version-gated external Pi extension and local broker.
  • ConnUI for the provider-neutral AppKit and SwiftUI notch surface.
  • ConnApp for opt-in Integration composition, setup, and migration controls.

Read the architecture decisions, domain model, and operations guide for the deeper contracts.

Privacy and safety

Conn reads bounded structured Session, Run, Activity, request, model, and status data needed for the visible supervision surface. The Pi broker is current-user local and validates bounded identities and acknowledgements. Conn does not poll transcript files, enable daemon remote control, or claim lifecycle ownership. Consequential actions remain bound to the exact Integration and Session identity.

Shared Desktop Mode uses an internal Codex Desktop switch that may change in a future release. Its setup is explicit, reversible, current-user-only, and documented in docs/shared-desktop-mode.md.

Building and testing

git clone https://github.com/architpai/conn.git
cd conn
./scripts/build-app.sh

swift run conn-codex-adapter-tests
swift run conn-pi-adapter-tests
swift run conn-domain-tests
swift run conn-app-core-tests
swift run conn-ui-tests
pnpm pi:typecheck
pnpm --filter @conn/pi-extension test
./scripts/test-inspect-release.sh

The built application is written to .build/conn-app/Conn.app. See INSTALL.md for installation and CONTRIBUTING.md for the full development workflow.

Built with Codex and GPT-5.6

Conn was created during OpenAI Build Week, from July 18–21, 2026. GPT-5.6 in Codex accelerated protocol research, Swift implementation, test generation, privacy hardening, release tooling, code review, and computer-driven UI verification.

The human-led decisions remained central: choosing the notch supervision problem, defining the interaction design, pivoting from a hook-based prototype to the managed App Server architecture, keeping Codex as the lifecycle owner, setting the privacy boundary, and accepting or rejecting each implementation and design tradeoff. The project was developed in phased Codex tasks with explicit plans, ADRs, deterministic tests, adversarial review, and live app checks.

Project status and roadmap

Version 0.2.1 is a continued alpha. It ships qualified Codex and external Pi Integrations on the provider-neutral Session model. Claude Code, OpenCode, and other harnesses are not supported. External Pi creation, approval, and question controls remain out of scope. Near-term priorities remain Developer ID signing and notarization, broader adapter qualification, universal macOS builds, and UI refinement through daily use. See CHANGELOG.md for release history.

Contributing and security

Contributions are welcome. Start with CONTRIBUTING.md. Please report vulnerabilities privately as described in SECURITY.md.

Conn is licensed under the Apache License 2.0. Third-party inspiration and generated protocol-schema provenance are recorded in ACKNOWLEDGEMENTS.md.

About

A native macOS notch companion for supervising Codex work.

Topics

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages