◎ Eye of God

Your personal, real-time command center for AI coding agents.

One dashboard over Codex + Oh My Pi — run any model live, watch agents think, track every provider's quota, and orchestrate a swarm of skill-backed agents. It owns no model; it borrows them all.

zero dependencies · python stdlib + vanilla JS · localhost-only · one command to launch

---

Why

You already have powerful coding-agent CLIs (Codex, Oh My Pi). What you don't have is a single pane of glass over all of them — your sessions, models, skills, agents, provider quota, and live work — that you can actually act from.

Eye of God is that pane. It reads your real ~/.codex data, streams live model runs through omp across ~65 models from every provider you're signed into, and turns your installed skills into a runnable agent swarm — all in a UI that feels like a modern desktop app.

It’s a generic supersystem wrapped around Oh My Pi — point it at your setup and it becomes your command center.

Highlights

• 🛰️ Live Agent Console — type a prompt, pick any of ~65 models (Anthropic / OpenAI / Google / xAI), and watch the agent think and answer in real time, streamed token-by-token with token count + cost + elapsed. Smooth incremental rendering (no full-page repaints).
• 🧠 Agent Swarm — a config-driven roster (Super Agent → Planner → Officers) plus one auto-generated specialist per installed skill. Click Run to launch any agent with its persona + skills + model. Works in the dashboard and as a copy-paste omp CLI command.
• 📊 Providers — live multi-account quota bars (5h / 7d limits, reset timers) for every provider, parsed from omp usage. A true eye over all your models.
• ⚡ Humanized Activity — raw OpenTelemetry log spans decoded into a friendly, interactive timeline with an activity chart, a composition donut, filter chips, and expand-to-raw.
• 🎛️ Settings UI — theme, default model, visible panels, working dirs — edited in-app and hot-reloaded.
• 🗂️ Boards — Kanban for Work / Personal / Ideas / Projects, editable and persisted.
• 🚀 Launcher — build & launch real Codex sessions (model / cwd / approval / sandbox / search) in a terminal, with a live command preview.
• 🔭 Plus Sessions, Models & Auth, Skills, Integrations (MCP + plugins), Memories, Logs — and a ⌘K command palette over everything.

Agent Swarm	Live Activity	Overview

Quick start

git clone https://github.com/astatineRS/eye-of-god.git
cd eye-of-god
./eye

That's it. No pip install, no npm install — just Python 3.11+. It seeds a config from the example, finds a free port, and opens http://127.0.0.1:7777. Press ⌘K / Ctrl+K for the command palette.

Make it a one-word global launcher:

ln -sf "$PWD/eye" ~/.local/bin/eye   # then just:  eye

Prereqs to unlock everything:

• Codex CLI (~/.codex) — powers the data panels (sessions, models, skills, logs, auth).
• Oh My Pi (omp) — powers the live Console, model catalog, and provider quota.
• Ollama (optional) — local model status.

Missing one? The dashboard still runs; those panels just show what's available.

How the agents work — dashboard vs CLI

An agent = a persona (system prompt) + a skill set + a model. Eye of God runs it through omp (which holds your provider auth) — so the dashboard never stores an API key.

	Best for	How
Dashboard Console	Quick framed asks, comparing models, watching work live with cost	Agents panel → Run in Console → type task
omp CLI	Deep multi-step work with full tools, approvals, sub-agents, saved sessions	Agents panel → Copy omp CLI → omp --skills "<glob>" "task"

Recommendation: use the Console to pick the right agent fast and triage; switch to the omp CLI (or Launcher → Codex) when the task needs to read/write files and run commands across many steps. Super Agent / Planner → routing & strategy; Officers → a domain; Specialists → a single skill.

Skills are discovered by omp from ~/.codex/skills (Eye of God links them onto omp's path). Define your own roster in dashboard.config.json under agents.

Architecture

flowchart LR
  B["Browser UI<br/>(vanilla JS)"] -->|HTTP + SSE| S["server.py<br/>(python stdlib)"]
  S -->|read-only| C["~/.codex<br/>sqlite · jsonl · config"]
  S -->|spawn + stream JSON| O["omp / codex"]
  O --> P["Anthropic · OpenAI<br/>Google · xAI · Ollama"]

Loading

• Backend (server.py, no deps): serves the UI, exposes /api/* reading real ~/.codex data (SQLite opened read-only), pushes a live snapshot over SSE, and runs agents by spawning omp -p --mode json …, normalizing its event stream (thinking / text / tool / usage) for the browser to poll.
• Frontend (web/): one STATE, one load(root) per panel, config-driven, SVG icons, smooth streaming.
• Everything is config — dashboard.config.json (theme, panels, agents, commands, integrations). Full guide in EYE_OF_GOD.md.

Customize

• In-app: the Settings panel (theme, default model, visible panels, dirs) → saved + hot-reloaded.
• By file: copy dashboard.config.example.json → dashboard.config.json and edit. Add agents, slash/CLI commands, integrations, boards.
• Extend a panel: add a provider in server.py + a GET_ROUTES entry + a PANELS loader in app.js. Three small edits.

Security

• Binds to 127.0.0.1 only.
• Secrets are never served — auth.json is read only to compute booleans; tokens/keys never reach the browser.
• SQLite opened read-only. The only writes are the dashboard's own data/ and (on explicit click) a backed-up config.toml model change.
• POST /api/run diagnostics are a small whitelist.

Roadmap

• Multi-turn console sessions (omp --continue)
• Agent-to-agent handoff (router auto-dispatch)
• One-click panel plugins
• Themes gallery

PRs welcome. Built on the shoulders of Codex and Oh My Pi.

License

MIT — see LICENSE.