curl -fsSL https://raw.githubusercontent.com/mkhuda/agent-ps/main/install.sh | sh

ps aux | grep claude gives you PIDs. It will not tell you which one is the session you are talking to, what model it is spending on, whether it has been idle for three days, or which rows are background daemons that outlived the terminal that started them. It also misses the other seven agents entirely.

agent-ps lists every session with its agent, model, working directory, uptime, idle time and disk footprint, then stops one process tree, every background helper, or the lot.

One file, no dependencies, reads local files only.

Install

One command

curl -fsSL https://raw.githubusercontent.com/mkhuda/agent-ps/main/install.sh | sh

Takes the latest release, checks it against the checksum published beside it, makes sure it runs, and puts the single file in ~/.local/bin. Set BIN_DIR to install elsewhere, or AGENT_PS_REF=v0.1.0 to pin a version.

With a package runner

From PyPI, which installs the program itself:

uvx agent-ps                  # run it once
pipx install agent-ps         # keep it, isolated
pip install agent-ps          # keep it, wherever pip points

From npm, which carries the same single executable and finds an interpreter for it:

npx @mkhuda/agent-ps
pnpm dlx @mkhuda/agent-ps
bunx @mkhuda/agent-ps

Either way the command it gives you is agent-ps.

Just the file

If you would rather not pipe a script into a shell, the executable is the whole program:

mkdir -p ~/.local/bin
curl -fsSL -o ~/.local/bin/agent-ps \
  https://github.com/mkhuda/agent-ps/releases/latest/download/agent-ps
chmod +x ~/.local/bin/agent-ps

Every release publishes a SHA256SUMS next to it.

From a clone

git clone https://github.com/mkhuda/agent-ps && cd agent-ps
./install.sh

Builds from the tree rather than downloading, so you install what you are looking at.

What it needs, and how to remove it

Python 3.8 or later and nothing from PyPI. Tested on 3.8, 3.9, 3.12 and 3.13. macOS and Linux; Windows is out because curses is not in its standard library. It shells out to ps, and on macOS to lsof, which is how a process is matched to a session everywhere except Claude Code.

rm -f ~/.local/bin/agent-ps

That is the whole uninstall. Nothing is written outside that directory, and agent-ps only ever reads the agents' own files.

Agents

Each agent keeps its own colour in the table, and the same colours label the legend above the keys.

Agent		Sessions in	Paired by	Reopened with
Claude Code		JSONL	the agent itself	claude --resume
Pi		JSONL	directory	pi --session
CommandCode		JSONL	directory	cmd --resume
Codex CLI		JSONL	directory	codex resume
OpenCode		SQLite	directory	opencode --session
Hermes		SQLite	directory	hermes --resume
GitHub Copilot		VS Code storage	no process	in the editor
Antigravity		SQLite per conversation	directory	agy --conversation

Only Claude Code records which process is running which session. Every other pairing is matched on working directory and shown as a guess. Tokens and cost appear in the detail panel.

Copilot is the exception: it runs inside the VS Code extension host, so it has no process of its own and nothing to stop. A chat counts as open while its workspace is open in the editor.

Usage

Run without arguments for the live table:

agent-ps

It refreshes every two seconds and stays out of the way until you act on something.

Those keys live at the bottom of the screen, under a line naming every agent on screen in its own colour and, on the right, what the table is sorted by:

That middle line is the key to the colours, names the sort order in words, and is where a note appears for a few seconds after you act on something.

Backspacing a filter down to nothing leaves filter mode, so the key bar comes back without reaching for escape.

Sorting

s cycles the sort column through agent, active, disk, cpu, mem, uptime, session and title. The sorted heading is marked in place, and the line above the keys names it in words:

PID     AGENT             SESSION         STATUS MODEL               UPTIME  ACTIVE    CPU   MEM    DISKv
31771   claude            benchmark       idle   sonnet-5            54m     54m ago   0.2%  51M    127M
-       claude            web-app         ended  sonnet-5            -       16d8h ago -     -      102M
-       claude            notes           ended  opus-5              -       2h22m ago -     -      84M

 agents  claude  codex  commandcode  copilot  hermes  opencode  pi        sorted by disk, high to low
 up/down move   enter details   k stop   / filter   s sort   e hide ended   ? keys   b background   p prune   S reverse   q quit

Keys drop from the right as the terminal narrows, q quit last of all. Press ? for the ones that never fit: home/end, space to pause, r to refresh now, y/n to answer a confirmation.

v means descending and ^ ascending; S flips it. Numbers start at the largest, names start at A, and rows with nothing in that column go to the end either way.

Sorting covers running and ended sessions together, since "which session is the biggest" does not care whether its process is still alive. Press e to include ended sessions, then s until DISK is marked.

The detail panel

Enter on a live session opens everything known about it, in four groups: what the session is, what it is doing to the machine, what it has cost, and the command line.

The disk total is broken down by what is holding it, since transcript and file history call for different answers.

Tokens and cost are here too, and Copilot reports the credits a turn spent:

    provider  opencode-free
      tokens  in 24,608  out 242  reasoning 190
       calls  2 api, 6 messages, 0 tool calls
        cost  $0.0000 estimated

A paired line appears when the session was matched to the process by working directory rather than reported, so you know before pressing k that the pairing is a guess.

A last error line appears when the newest turn failed and nothing has succeeded since: a rate limit, an auth failure, a server error. Claude Code, Codex CLI and OpenCode report this; going through a custom ANTHROPIC_BASE_URL or similar proxy, a failure there is invisible to the transcript and so invisible here too, and the session just goes quiet.

Enter on an ended session asks before reopening it in a new terminal tab, since the session can start spending the moment it opens. Copilot chats reopen in the editor.

The advisory line

A line above the keys points out whatever is worth a look: a live session that just hit an API error, background helpers left running, sessions untouched for a day, or ended sessions you could resume. It names the key that acts on it where there is one, and shows one thing at a time.

Scripting

agent-ps list                     # print the table and exit
agent-ps list --all               # include ended sessions
agent-ps list --json              # machine readable, with idle seconds
agent-ps list --limit 100         # how many ended sessions (default 40)
agent-ps list --filter benchmark  # same match as the / key
agent-ps prune                    # what ended sessions left, minus the conversation
agent-ps prune --older-than 30d   # only sessions idle that long (default 7d)
agent-ps prune --apply            # actually remove it
agent-ps agents                   # which agents were found, and where
agent-ps --agent codex list       # one agent, or a comma separated list
agent-ps stop 32244               # stop one process tree
agent-ps stop 32244 --dry-run     # show what would be stopped
agent-ps stop-background          # stop daemons, warm spares, and servers
agent-ps stop-background --dry-run
agent-ps resume <session>         # a unique id prefix is enough
agent-ps resume <session> --print # print the command instead of running it
agent-ps --version

Piping works without a subcommand: with stdout not a terminal, agent-ps prints the table and exits.

Reclaiming space

A session writes more than its conversation. Claude Code alone keeps subagent transcripts, file history for /rewind, task records and a session environment, and on this machine those came to a fifth of everything on disk.

agent-ps prune

Reports and removes nothing. It lists what ended sessions left behind, grouped by what it is and which sessions hold the most, and --apply removes it.

In the table, p does the same for the row you are on, and the detail panel marks which parts of a session those are. On a session that is still running it refuses and says why rather than going quiet.

Three things it will not do. It never touches a transcript, so every session stays resumable. It never touches a session with a running process, whatever its age, because file history is what /rewind reaches for. And it never touches OpenCode, Copilot or Antigravity, whose sessions are rows in a database the agent may have open: agent-ps reads, and does not write, the stores it did not create.

What each column means

UPTIME and ACTIVE often disagree: a process can be five days old and have answered a minute ago. UPTIME comes from the process table, ACTIVE from the log, which is appended on every turn.

Each agent has its own colour, and the legend above the keys is the key to it. Terminals without colour fall back to plain text.

A PID marked with a question mark

Only Claude Code records which process is running which session, in <config dir>/sessions/<pid>.json. For every other agent the two have to be matched on working directory, and that cannot tell apart two sessions of the same agent started in the same folder.

So an inferred pairing is shown as one:

32244   claude    form-guardian   idle   opus-5        ...
9666?   codex     agent-ps        idle   gpt-5.6-luna  ...

9666? means the process is certain and the session beside it is a guess. The guess is the most recently active session in that directory, and where several processes of one agent share a directory they are paired newest with newest, so the process that started four minutes ago gets the session that has been active for four minutes rather than the one from an hour ago.

It is still a guess, so the stop confirmation names the directory and says when the pairing was inferred:

Stop 1 process(es) in ~/projects/agent-ps, session matched by directory?

Where no session matches at all, the row still appears with whatever the process itself can answer. Nothing is invented to fill the gap.

How it works

The parts that are not obvious have their own page: where each agent keeps its sessions, how busy and idle are decided per agent, which processes count as background helpers, how a shim is told apart from a real session, and what the DISK column is adding up.

Stopping behaviour

Processes are stopped depth first, children before parents, so a supervisor does not restart a worker you just killed. Each process gets SIGTERM, one second to exit, then SIGKILL if it is still there.

agent-ps never lists or stops itself, or the shell that launched it, so running it from inside an agent session is safe.

Exit status is zero when everything asked for was stopped, and one when something survived, which makes it usable in scripts.

When something is missing

Build

The executable is a zipapp, built from the package by the standard library:

./build.sh

That writes a single agent-ps file with no dependencies, which is what the curl install fetches. The build is reproducible: the same source always produces the same bytes, so the committed executable can be checked against the tree.

9f5c42aeb1fe52ed69781731bcfb6c945b028e81f7067122ec925e2451f1a0ca  agent-ps

Adding an agent takes one class and one line in the registry. See CONTRIBUTING.md.

License

MIT. See LICENSE.