Local Command Center

Open-source native macOS continuity workspace for coordinating local Codex and Claude Code sessions. The public build is intentionally local-only: provider credentials, raw transcripts, local session identifiers, and machine-specific paths never enter the repository or the portable continuity package.

A native, local-only macOS workspace for organizing coding-agent chats and routing work through the authenticated Codex and Claude subscription CLIs.

The local milestone intentionally avoids Electron, WebViews, a local HTTP server, background model pools, and cloud synchronization. One Swift process owns the UI and local SQLite state; provider processes exist only while work is running.

Open-source release boundary

This repository contains the public application source and reproducible tests. It does not broker consumer subscriptions, create web Projects, scrape browser state, or synchronize provider-owned transcript stores. Local transport and provider session lineage remain protected runtime state on the host machine.

Local history federation milestone

• Native project, pinned, inbox, and recent chat organization.
• Local SQLite persistence with a versioned schema and task-scoped skill snapshots.
• Automatic read-only indexing of locally materialized Codex and Claude Code sessions, grouped by project with accessible provider badges.
• Lazy recent transcript windows (50 messages / 2 MiB) from provider-owned JSONL; provider files remain authoritative and are never rewritten.
• UUID-safe continuation through the already authenticated Codex and Claude subscription CLIs after one exact project-folder approval and a fresh, bounded source revalidation before each linked launch.
• Launch, foreground, provider-exit, manual, and coalesced periodic refresh.
• Optional deterministic metadata/backlink projection to a registered Obsidian vault under an app-owned Command Center/ subtree.
• Direct, pickup-swarm, and paired-review dispatch modes.
• Visible skills, permission mode, provider, queue, and memory pressure.
• Adaptive concurrency that preserves headroom without slowing active work.

Build and run

./script/build_and_run.sh

If a managed execution environment blocks SwiftPM's own nested sandbox, use COMMAND_CENTER_DISABLE_SWIFTPM_SANDBOX=1 ./script/build_and_run.sh; the built application's runtime and privacy model are unchanged.

Useful gates:

swift test
./script/build_and_run.sh --verify
./script/build_and_run.sh --profile

Operational state remains under ~/Library/Application Support/Local Command Center/. No provider credentials are copied into the application database.

First use

1. Open dist/Command Center.app.
2. Select Local history to browse locally materialized Codex and Claude Code sessions. Select one and use Approve project and continue… once per exact project root; archived Codex sessions remain visibly labeled.
3. Or select Add project… and choose one repository folder, then create a new task. Choose Codex or Claude, keep Read only for questions and reviews, and use Workspace write only for changes you want the provider to make inside that selected folder.
4. Choose Direct, Pickup swarm, or Paired review, select any installed skills, and press ⌘Return.
5. Watch the persistent goal card and bottom status bar; ⌘. stops the selected app-owned provider process.

The app renders only bounded transcript windows to keep long-running histories from exhausting memory. External transcript bodies are not duplicated into SQLite or Obsidian. After the explicit in-app disclosure, the metadata-only Obsidian projection contains opaque IDs, provider-derived task titles (which may come from a first prompt), project folder labels/backlinks, status, and timestamps. Obsidian Sync or plugins may export that metadata. Command Center never reads or modifies existing notes.

Consumer-only ChatGPT and Claude.ai web chats are BLOCKED_PROVIDER_CAPABILITY unless an official local read-and-resume surface materializes them. Command Center intentionally does not scrape desktop caches, browser data, or tokens.

Automatically launched paired reviews stay classified as background work for admission priority and display, while their provider prompt remains a focused direct code review. Final stdout/stderr bytes are drained before a provider turn is marked complete, so fast-exiting CLIs cannot silently lose their last event.