Codex Control Center

Local observability and approval-gated control for Codex.

No API key required for local observation. The dashboard reads local Codex metadata, local session JSONL files, skills/plugins metadata, schedules, and optional local OTel events. It never reads or stores auth.json, and it binds to 127.0.0.1 for local-only use.

Public Links

• Website case study: ai.joaoqueiros.com/systems/codex-control-center
• Blog guide and build prompt: Build Your Own Local AI Agent Control Center
• Interactive demo: jqaisystems.github.io/codex-control-center/demo
• Video hub: jqaisystems.github.io/codex-control-center/demo/videos.html
• YouTube videos: Official Launch, Step-by-Step Tutorial, Full Browser Walkthrough
• Release: v0.1.0

Screenshots

These screenshots use fake demo data only.

Interactive Demo

A public-safe self-playing HTML demo is available at demo/index.html. It uses fake dashboard data and does not connect to Codex, read local files, call OpenAI, publish anything, or reveal private paths. The included demo music is owner-created by JQAI Systems.

Open the hosted demo: https://jqaisystems.github.io/codex-control-center/demo/

Watch embedded videos: https://jqaisystems.github.io/codex-control-center/demo/videos.html

Build Your Own

Use the public-safe prompt in build-your-own-codex-control-center-prompt.md to build a local-first AI agent control center inspired by this project.

The prompt can be adapted for Codex, Claude Code, Cursor, or other coding LLMs. Ask the assistant to adjust CLI commands, session paths, event formats, and sandbox controls for the tool you use while preserving the same privacy rules.

Open the website version: https://jqaisystems.github.io/codex-control-center/demo/build-your-own.html

Watch

• Official Launch: 60-second overview of Codex Control Center v0.1.0.
• Step-by-Step Tutorial: safe setup and usage walkthrough.
• Full Browser Walkthrough: browser tour of the dashboard, tasks, reports, skills, and publish checks.

Companion video scripts, storyboards, and HyperFrames source projects live in docs/video/ for the launch presentation, practical tutorial, and browser walkthrough. Rendered MP4 drafts stay local in video-output/ until they are reviewed and intentionally uploaded elsewhere.

What It Shows

• System Health: Codex CLI status, local session coverage, last sync, OTel state, database label, and privacy posture.
• Usage Remaining: best-effort local usage metadata parsed from Codex token_count events with payload.rate_limits.
• Today Activity: sessions, token totals, tool counts, recent sessions, and top tools from local metadata.
• Tasks: approval-gated task queue for dashboard-launched codex exec --json runs.
• Schedules: local schedule definitions that materialize due work into approval-gated tasks only.
• Results: searchable run history with safe output summaries, durations, event counts, tool counts, exit codes, and sanitized failure reasons.
• Skills: discovered local Codex skills and plugin-provided skills.

Modes

• Observe Mode: works without an API key and does not call OpenAI.
• Control Mode: launches approved tasks through your installed codex exec --json --ephemeral. If Codex is signed in with ChatGPT, this does not require an API key.

Quick Start

Tested on Windows 11. Windows is the primary supported setup for V1; macOS and Linux may work but are not fully tested yet.

Windows launcher:

cd codex-control-center
.\start-control-center.ps1

The launcher creates a local Python virtual environment if needed, installs missing dependencies, builds the React UI when ui/dist is missing, starts the backend on 127.0.0.1:8765, waits for /api/health, and opens the browser.

Useful launcher options:

.\start-control-center.ps1 -NoBrowser
.\start-control-center.ps1 -Rebuild
.\start-control-center.ps1 -Install
.\start-control-center.ps1 -Foreground
.\start-control-center.ps1 -StopExisting

Manual setup:

cd codex-control-center
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
cd ui
npm install
npm run build
cd ..
python -m backend

Open http://127.0.0.1:8765.

For frontend development:

cd codex-control-center
python -m backend
cd ui
npm run dev

Open http://127.0.0.1:5173.

Privacy Defaults

• Metadata-only ingestion is enabled by default.
• Prompt text and assistant output from local Codex sessions are not stored.
• Raw local paths are redacted to project basename plus a stable local hash.
• The backend never reads ~/.codex/auth.json.
• The dashboard does not call OpenAI directly.
• Mutating endpoints are loopback-only unless CCC_CONTROL_TOKEN is set.
• Task execution output is reduced to redacted summaries and metadata.

Task Safety

• New tasks start as awaiting_approval.
• Default sandbox is read-only.
• workspace-write must be selected explicitly.
• danger-full-access is blocked in V1.
• Schedules only create approval-gated tasks; they do not auto-run Codex.
• Emergency stop only targets dashboard-launched child processes.

For starter prompts and a safe task-writing flow, see docs/task_templates.md.

Video Kit

Public-safe launch and tutorial video scripts live in docs/video/, including the launch presentation, practical tutorial, and browser walkthrough. Rendered videos should stay local in video-output/ unless they are intentionally reviewed and uploaded.

Optional OTel

Codex OTel is optional. The dashboard works from local session files without it. To see a local-only config block:

.\scripts\setup_otel.ps1

Review the output before adding it to your Codex config.

Public Publishing Safety

This folder is the public candidate. Do not publish raw Codex session files, databases, .env, logs, private screenshots, local paths, prompts, or account data.

Run before committing or pushing:

python scripts/public_safety_scan.py .
git status --short
git diff