This aitop: a live, read-only monitor of your AI coding agents and their contexts — Claude Code, Codex CLI, Cursor, cursor-agent CLI, opencode — one card per session, showing what it's doing, which model is driving it, how full its context window is, and its token burn, all in one terminal pane. System resources (CPU/MEM/NET) are a condensed footnote at the bottom, not the headline. If you were looking for a pure cost/token tracker, that's a different aitop (bugkill3r/aitop or samusgray/aitop) — this one is a context board for your agents, btop-styled, not a spend dashboard.
- A card per agent session, always expanded. Tool identity (border color) and, when the adapter knows it, the model actually driving that session —
claude code (opus 4.8),opencode (deepseek v4 flash free)— a title synthesized from the session's own first request (or Claude Code's own auto-generated title, when there is one), the state badge, the last action taken (tool call or message, word-wrapped), aContext: [bar] 234k/1000k (23%)line — a real reading, never a percentage guessed from nothing — and its 5h/7d rate-limit + process detail underneath, across every Claude Code, Codex, Cursor, cursor-agent, and opencode session running on your machine right now. - Per-session, not per-tool. Two Claude Code sessions running side by side get their own token counts and context bars — they never mix, and a card for a dead session never lingers as if it were still running.
- Multi-tool from day one. Claude Code, Codex CLI, Cursor, cursor-agent CLI, and opencode — Cursor and cursor-agent in particular aren't covered by any other "aitop"-named project as of this writing.
- Resources, condensed. A short footer shows real aggregate CPU/MEM/NET (via gopsutil); the MEM bar is segmented to show how much of that is attributable to your AI-tool processes specifically — present, but deliberately secondary to the cards above it.
- Read-only. No approve/reply/merge actions (that's agent-dashboard's job, and it requires tmux — aitop doesn't). aitop only ever observes.
- Honest about gaps. Missing data renders as
—, never a fabricated0. Cursor has no local cost data and aitop says so instead of guessing; the same goes for any field a given adapter hasn't populated yet.
curl -fsSL https://raw.githubusercontent.com/grippado/aitop/main/install.sh | shgo install github.com/grippado/aitop/cmd/aitop@latestHomebrew tap (brew install grippado/aitop/aitop) is planned but not live yet — it needs a separate grippado/homebrew-aitop repo this release doesn't publish to.
aitop # live TUI, 2s refresh
aitop --once # one text snapshot, then exit
aitop --once --json # machine-readable snapshot
aitop --refresh 5s # override the tick interval
aitop --demo # synthetic agent cards, no real tool required| Key | Action |
|---|---|
q / Ctrl+C |
quit |
j/k, arrows |
move selection |
f |
filter cards by tool |
o |
cycle sort column (context / tokens / age / tool) |
v |
toggle list/grid card layout |
space |
pause/resume refresh |
r |
force refresh |
? |
help |
Every existing "aitop"-named project (and the excellent but tmux-bound agent-dashboard) either tracks cost/tokens as the headline, requires tmux, or covers Claude Code alone. None of them present agent sessions as a live, glanceable board the way a mutirão of open threads deserves, and none of them support Cursor or cursor-agent. aitop's whole reason to exist is that gap: a real-time context board for the agents actually running on your machine, with resource usage as a footnote instead of the main event.
| Tool | Process/CPU/MEM | Sessions | Title / last action | Model | Context% / tokens | Rate limits |
|---|---|---|---|---|---|---|
| Claude Code | ✅ | ✅ | ✅ (own auto-title + transcript) | ✅ (own transcript) | ✅ (per-session, own transcript) | ✅ (5h/7d, via ccstatusline's cache) |
| Codex CLI | ✅ | ✅ | ✅ (synthesized from first request + transcript) | ✅ (own rollout's turn_context) | ✅ (per-session; Codex's own rollout gives the true model context window, no guessing) | Not available (no local rate-limit source) |
| Cursor (IDE) | ✅ (Cursor's own telemetry) | ✅ | ✅ (real title + last message/tool call, from Cursor's own state.vscdb) |
Not available | Tokens: ✅ (per-composer, same source) · Context%: not available (no window-size reading found) | Not available (proprietary/cloud-side) |
| cursor-agent (CLI) | ✅ | ✅ | ✅ (first request + own transcript) | ✅ (currently selected model, from cursor-agent's own config) | Not available (no per-turn usage in its transcript format) | Not available |
| opencode | ✅ | ✅ | ✅ (opencode's own session title) | ✅ (own models catalog cache gives the real display name) | ✅ (per-session, straight from opencode's own SQLite state; real context window from its models catalog) | Not available (its own cost figure is lifetime-cumulative, not day/month-scoped — see the adapter's own doc comment) |
| Other AI CLIs (aider, windsurf, ...) | Best-effort via process-name match | — | — | — | — | — |
Cost-in-dollars was dropped from the expanded card view: on this machine's real data the cost-day/cost-month files aitop originally read haven't been written in weeks, so it never showed anything but a fake $0.00. Fields still unpopulated by any adapter (git branch/dirty state) render as — on real data today; --demo shows what the card looks like once those land.
v2 roadmap: dedicated adapters for aider/windsurf, an external plugin mechanism, branch/dirty detection, a local trend history, and community-contributed themes (see CONTRIBUTING.md).
MIT