Amir Fish just shipped v4.6.0 of the Claude Command Center, a macOS-only dashboard that attaches to every AI coding agent running on your machine—terminal sessions, headless processes, and ones spawned directly from the UI. The latest update (June 3) fixed CPU pinning issues during idle states, cut group-chat load time by roughly 40x, and addressed multi-MB screenshot stalls in Codex sessions. If you're juggling multiple Claude Code workflows across projects, this tool deserves your attention.
How It Works
CCC treats Claude's on-disk state as the source of truth rather than wrapping execution like most orchestration tools. The server reads ~/.claude/projects/*.jsonl transcript files, ~/.claude/sessions/
Multi-Engine Support
The project started Claude Code-first but has expanded to handle Codex (via OpenAI), Cursor (via cursor-agent CLI), and Antigravity (Google DeepMind, added in v4.0.0 on May 21). Spawn-from-dashboard works across all four engines, but transcript ingestion and UX parity vary. Claude Code gets first-class JSONL parsing from ~/.claude/projects/, while Codex and Cursor support is marked "partial" in the project's own engine matrix. Antigravity ingests JSONL transcripts from ~/.gemini/antigravity/brain/ with auto-detected model info.
Key Features
The kanban board classifies sessions into columns (Backlog → Planning → Working → Review → In Testing → Verified/Inactive/Archived) based on session state—live processes, commit presence, push status—but you can drag cards to override. GitHub integration lets you start a session from an issue with one click (auto-labels claude-in-progress and self-assigns), verify it closes the issue with a commit-SHA comment, or drag to Archived to close as "not planned." There's also Vercel deploy polling that spawns /fix-deploy sessions on new production ERRORs, deduplicated by commit SHA.
Installation Options
Getting started requires macOS, Python 3, and Claude Code installed. The fastest path: curl -fsSL https://raw.githubusercontent.com/amirfish1/claude-command-center/main/scripts/install.sh | CCC_FROM=readme bash. Homebrew users can tap amirfish1/ccc and brew install ccc. For the no-install experience, try the read-only demo at ccc.amirfish.ai/demo with seeded fake data. The server runs on localhost:8090 by default (configurable via PORT env var), and ./run.sh --install-service writes a launchd plist to ~/Library/LaunchAgents/ for background startup.
Orchestration Skill
For AI-to-AI coordination, CCC ships a ccc-orchestration skill that lets one Claude session spawn, inject into, and synchronously query sibling sessions over plain HTTP. The server writes its base URL to ~/.claude/command-center/port.txt so the skill auto-discovers the running instance. Spawn calls pass repo_path plus optional engine: "claude" | "codex" | "cursor" | "antigravity", with omitted values falling back to dashboard defaults. This enables persistent peer sessions—a marketing assistant, a deploy babysitter—that survive past the current turn and appear on the kanban.
Under the Hood
Two files: server.py (stdlib-only HTTP server) and index.html (vanilla JS, no framework). The architecture intentionally avoids background workers—every API request scans Claude's session directories, merges sidecar state, enriches with cached GitHub data, and returns a flat list. On first launch, the server copies hooks/post-tool-use.py and hooks/stop.py to ~/.claude/command-center/hooks/, then registers them in settings.json so every tool invocation writes tiny JSON files under live-state/. The client classifies into kanban columns using rules like "has_push → Review."
Key Takeaways
- CCC v4.6.0 fixes CPU pinning and performance regressions from earlier releases
- Claude Code support is first-class; Codex/Cursor/Antigravity vary in maturity
- Treats existing on-disk state as source of truth—no vendor lock-in to the dashboard itself
- GitHub + Vercel integrations automate issue tracking and deploy fix workflows
- One-person project by @amirfish1—contributions welcome, roadmap tracked in issues