From: Matt Brassey <matt@brassey.io>
To: submit@bugs.debian.org
Subject: ITP: agtop -- terminal UI for monitoring AI coding agents
Package: wnpp
Severity: wishlist
Owner: Matt Brassey <matt@brassey.io>
X-Debbugs-CC: debian-devel@lists.debian.org
* Package name : agtop
Version : 2.4.2
Upstream Author : Matt Brassey <matt@brassey.io>
* URL : https://github.com/MBrassey/agtop
* License : MIT
Programming Lang: Rust
Description : terminal UI for monitoring AI coding agents
agtop is a top-like process monitor specialised for AI coding
agents. It detects Claude Code, OpenAI Codex, Block Goose, Aider,
Google Gemini, Cursor, Continue, Opencode, Copilot CLI, and other
agent CLIs out of the box, and reads each one's session transcript
on disk to enrich its row with the current tool, model, in-flight
subagents, cumulative token usage, estimated dollar cost,
context-window fill, and loaded skills / plugins.
The interactive TUI groups agents by project, supports filter and
sort, opens a detail popup with a live transcript preview, and
exposes machine-readable JSON output for piping into dashboards.
A `--watch` mode emits one summary line per tick with optional
CPU% / tokens-per-minute thresholds for unattended monitoring.
No network access is performed at runtime — the cost estimator
ships a snapshot of LiteLLM's community model price registry and
can be overridden with --prices PATH.
I plan to maintain this package in Debian. Upstream is the same
maintainer. The package builds with dh-cargo against Debian's
existing librust-*-dev archive crates; no new dependencies need
introducing for it to land. Packaging lives at
https://github.com/MBrassey/agtop/tree/main/debian.
Initial sponsor request will go through mentors.debian.net under
the rust-team umbrella.
Rationale for inclusion: AI coding agents (Claude Code, Codex,
Aider, etc.) are increasingly common on developer workstations,
yet `top`, `htop`, `btop` show only opaque `node` / `python` /
`claude` rows with no insight into what those agents are actually
doing. agtop fills the gap with a small, fast Rust binary that
reads the on-disk transcripts each agent already keeps and
correlates them with /proc data — useful for the increasing
number of Debian users who run several of these agents in parallel
and need to monitor cost, context-fill, and runaway behaviour.