top
for Claude Code. Every live session, the model it is on, how much context it has burned — and, unlike anything else, the subagents it spawned.
One file, no dependencies, Python 3.9+. Runs on macOS, Linux and Windows.
The short ambient loop below is the same program, idling:
WORKER MODEL CTX TOKENS TREND FLOW IDLE TASK
NEAR LIMIT
demo-a1 opus-5 85% 170k +4k ..:-=+#+=-.. 12s refactor the parser
PARKED + COSTLY
demo-b2 opus-5 61% 122k +2k .............. 4h10m audit the build scripts
WORKING NOW
demo-c3 fable-5 22% 44k +22k ...:=+*#+ 3s add integration tests
STARTING
demo-d4 - - - - -
QUIET (4) demo-e5 . demo-f6 . demo-g7 . demo-h8
8 worker(s) | 1.2M held | 61k last 8 turns | 1 near limit | fable-5, opus-5
SUBAGENTS
STATE AGENT MODEL CTX IDLE TASK
working Explore/a812a opus-5 66k/200k 2s survey the config s
idle adaffaba4b sonnet-5 484k/1M 1h22m draft the migration notes
2 subagent(s), 1 working
INFRA ollama:11434 up qwen2.5-coder:14b (9.2 GB) litellm:4000 up openwebui:8080 DOWN
Sessions are grouped by what it costs to ignore them, not by size: NEAR LIMIT
is about to stop working, PARKED + COSTLY
bills its whole context on the next turn, and everything quiet collapses to a single line.
TREND
is context added over the session's last few turns. It is an amount and
not a sparkline because context inside a session only ever rises — it drops
solely on /compact
— so a shape would draw the same ramp on every row. CTX
says how full a session is; TREND
says how fast it is filling, which is how
demo-c3
above shows as the one to watch at 22%. It is read back out of the
transcript rather than accumulated while roost runs, so it is there on the first
frame and under --once
. That is also what separates it from FLOW
alongside
it: FLOW
sparks throughput sampled since roost started and begins empty,
TREND
is an amount already in the file when roost opens it.
The last line totals the fleet. Context held is what a sweep would reclaim.
Homebrew (macOS and Linux):
brew install gmhoward9289-ops/tap/roost
Debian and Ubuntu, from the repo — this also gets you apt upgrade
:
curl -fsSL https://gmhoward9289-ops.github.io/roost/roost-archive-keyring.asc | sudo gpg --dearmor -o /usr/share/keyrings/roost-archive-keyring.gpg
echo "deb [signed-by=/usr/share/keyrings/roost-archive-keyring.gpg] https://gmhoward9289-ops.github.io/roost stable main" | sudo tee /etc/apt/sources.list.d/roost.list
sudo apt update
sudo apt install roost
Or, without adding a repo — grab roost_<version>_all.deb
from the latest release and install that file directly, e.g.:
sudo apt install ./roost_<version>_all.deb
Both resolve python3
exactly as a normal repo install would. The repo is
signed with a dedicated key (not tied to any personal identity); its public
half is packaging/apt/pubkey.asc
in this repo, published unchanged as
roost-archive-keyring.asc
above.
With pipx:
pipx install roost-top
PyPI holds the bare name roost
in reserve — a prior project's name, retained
after deletion — so the package is roost-top
; the command it installs is
plain roost
. Installing straight from the repo skips the index entirely:
pipx install git+https://github.com/gmhoward9289-ops/roost
With npm — the one channel that gives Windows a real roost
command:
npm install -g roost-top
Or run it without installing: npx roost-top
. The npm package is a wrapper, not
a port: it ships the same roost.py
and finds a Python to run it with (py -3
first on Windows, python3
elsewhere, 3.9 or newer either way). Python still has
to be on PATH
— npm delivers the script and puts roost
on PATH
, it does not
bring an interpreter. As on PyPI, the bare name roost
was already taken, so the
package is roost-top
and the command is plain roost
.
Or just take the file. It is one script, stdlib only, no dependencies:
curl -o roost https://raw.githubusercontent.com/gmhoward9289-ops/roost/main/roost.py
chmod +x roost && ./roost
With winget:
winget install gmhoward9289-ops.roost
That installs a frozen Windows executable — no Python required. It's built and
attached to every tagged release alongside a roost-<version>-windows-x64.zip
on the Releases page, if you'd rather grab the zip directly.
Or, without any of that: save roost.py
and run it — .PY
is in PATHEXT
,
so roost.py
works from anywhere on PATH
. That route and the npm one above
both still need a Python interpreter on PATH
; the winget install is the one that doesn't, since it ships a frozen exe.
The man page (man roost
) ships with the Homebrew and .deb
installs. A pipx
install puts it under the venv's own share/man
, which is not on the default
MANPATH
; read it in place with
man "$(pipx environment --value PIPX_LOCAL_VENVS)/roost-top/share/man/man1/roost.1"
.
roost live, refreshing every second
roost -w 5 slower refresh
roost -1 one frame, then exit
roost --json joined records, for piping
While running: space
refresh now · a
advice panel · s
subagents panel · m
local models panel · u
usage panel · g
gateway panel · r
remote panel · h
or ?
what am I looking at · i
arm interactive · q
quit
The FLOW
column is a sparkline of each session's recent token throughput —
context growth per refresh, normalised to its own busiest moment, newest at the
right. A .
is a sample with no flow; the ramp :-=+*#
is increasing activity. It is ASCII on purpose (block-drawing characters mojibake in the Windows console) and starts empty: history begins when roost starts, nothing persists.
The subagent CTX
column reads 48k/200k
— tokens loaded over the window.
The window is inferred (the smallest standard tier the observed usage fits in),
because nothing in a transcript records which window the session was opened
with. The AGENT
column shows the agent's type (Explore/a812a
) once the parent has recorded it — which only happens when the agent finishes, so a still-running agent shows its hex id.
Experimental, and off by default.i
arms interactive mode — the cursor,x
,y
, and theEXPERIMENTAL
marker in the top-right corner all come alive together, and it means it:x
ends a real process. Reading the dashboard has never been the risky half;i
is the one key that draws the line. Press it again to disarm — the cursor drops andx
/y
/j
/k
stop responding until you press it once more.
Once armed, j
/k
(or the arrow keys) raise a cursor. Raising it expands the
QUIET
group, because a session idle for hours is exactly what a sweep is looking for and it is unreachable while collapsed.
| key | does |
|---|---|
i |
|
| arm or disarm interactive mode | |
j k ↓ ↑ |
|
| move the cursor (interactive mode only) | |
x |
|
stop the selected session — confirms first, and only y proceeds |
|
y |
|
copy its sessionId, for claude --resume <id> |
|
esc |
|
drop the cursor, re-collapse QUIET |
Start already armed with roost --interactive
if you know you'll be acting on a session right away.
x
ends a process. It does not compact, save, or otherwise negotiate with the
session — there is no local control channel into a running Claude Code
session, so nothing gentler is available from outside it. On Unix that is a
SIGTERM
and the session exits on its own terms; on Windows there is no
cross-process equivalent, so it is a TerminateProcess
hard kill. Transcripts are written a turn at a time, so at most an in-flight turn is lost.
Both keys act on the row object that was on screen when you pressed them, never on an index re-resolved afterwards. Rows reorder between frames as sessions go quiet, and an index that outlived its frame would eventually stop the wrong one.
roost refuses to stop its own process or its parent — run it from inside the session it is pointed at and the cursor can land on the row that owns your terminal.
Only one panel is open at a time: a
, s
, m
, u
, g
, r
, and h
flip between
ADVICE, SUBAGENTS, LOCAL MODELS, USAGE, GATEWAY, REMOTE, and HELP rather than stacking. With two dozen sessions
on screen a stacked second panel lands below the bottom of the terminal, which
is indistinguishable from the key not working. For the same reason the frame
now says ... N more line(s) below
instead of quietly truncating.
h
or ?
opens a HELP panel — not a keybinding reference (the footer hint already lists the keys), but a one-line-each rundown of what each screen on the display means: INFRA, WORKERS, SUBAGENTS, ADVICE, LOCAL MODELS, USAGE, GATEWAY, REMOTE. roost is small enough that this is the whole manual.
The INFRA line only ever shows what Ollama currently has resident in VRAM —
ollama:11434 up qwen-coder-16k:latest (5.5 GB)
— because it reads
/api/ps
. A model that is installed but idle drops out of that line entirely, which reads as "roost doesn't see it" rather than "it isn't loaded right now."
Press m
for the full picture: every model ollama list
knows about, each
row showing disk size, residency, VRAM when loaded, and how long until Ollama
unloads it. It reads /api/tags
merged with /api/ps
; if Ollama isn't running, the panel says so instead of showing nothing.
u
opens the USAGE panel: tokens per day per model over the last week, tallied from the transcripts on disk (input + output from each assistant turn; cache reads are excluded on purpose — they are billed and limited differently, and counting them would swamp the number with re-reads of unchanged context).
USAGE observed transcript tokens (input+output) -- an estimate, not the Anthropic meter
2026-08-02 2.9M opus-5 1.6M, sonnet-5 965k, fable-5 253k <- today
2026-08-01 9.1M opus-5 4.5M, fable-5 3.3M, sonnet-5 984k
2026-07-29 1.9M gemma4-32k (local) 5.3M, opus-5 1.9M
today 2.9M | 7d 24.8M cloud / 60.0M budget (41%)
Two honesty rules bake into this panel. First, it is an estimate, not the Anthropic meter — there is no local file or API that records a plan's real rate-limit balance, so the budget is a number you set yourself:
export ROOST_WEEKLY_BUDGET=60M # or 850k, or a plain token count
Run /usage
inside Claude Code once, pick a number that matches what it shows,
and the panel tracks your burn against it from then on. Unset, the tallies show
without the budget fraction. Second, local models are free — anything
without a claude-
model name (Ollama via LiteLLM writes transcripts too) is
flagged (local)
in the breakdown and excluded from the cloud totals and the budget math.
The first u
scans a week of transcripts and can for a moment; after that only appended bytes are read, so keeping the panel open costs almost nothing. Day boundaries are UTC, because transcript timestamps are.
g
opens the GATEWAY panel: whether the LiteLLM proxy answers on
127.0.0.1:4000
, plus one row per batch-extraction run under the batch
directory (ROOST_BATCH_DIR
, default ~/litellm-server/batch
):
GATEWAY
litellm up (127.0.0.1:4000) last request 42s ago 3 req/min
jobs queue: inbox 0 running 1 done 12 failed 0
BATCH RUN MODEL DONE/TOTAL FAIL RATE ETA LAST
results-laneB-derived gemma4-32k 121/300 2 64/hr 2h48m 35s ago
results-laneA-derived qwen-coder-16k 5/5 0 - done 3d ago
The proxy itself is asked nothing beyond "are you up" — a DB-less LiteLLM
keeps no request history (every activity endpoint 400s), so progress is
derived from the batch pipeline's own output files: one JSON per finished
item means done/total, failure count, write rate, and ETA all fall out of a
directory listing. Runs whose extract.py
wrote a _run.json
breadcrumb show
their model and worklist; older runs still appear, just with less detail.
Green rows are actively writing; the last-request/req-per-min figures are a
best-effort read of proxy.log
's tail and disappear rather than guess when
the log doesn't parse. The jobs queue
line counts the file-based job queue's
dirs (JOBS_ROOT
, default ~/jobs
) if present.
r
opens the REMOTE panel — other machines' roost, over ssh:
REMOTE
HOST WORKERS WORKING RESIDENT MODELS BATCH JOBS AGE
hyrule 8 2 - - in 0 run 1 fail 0 12s
Hosts come from ROOST_REMOTES
(comma-separated ssh aliases, default
hyrule
) — from the environment only, never from file contents. Each host is
fetched with ssh <host> roost --json
on a background thread and cached: an
unreachable host (a closed laptop lid) keeps its last good row with the AGE
column saying how old it is, instead of hanging the display. Only the very
first fetch per host blocks, which is what makes roost --remote -1
useful.
ROOST_REMOTE_CMD
overrides the remote command if roost lives somewhere the non-login ssh PATH can't see.
Every session stopped with x
appends one JSON line to
~/.claude/logs/roost.jsonl
— same shape and the same 5000-line cap as the hook logs beside it:
{"ts":"2026-07-31T00:22:57-0400","action":"stop","ok":true,"host":"COOPER",
"name":"models-ca","pid":4321,"session_id":"abc-123","model":"claude-opus-5",
"ctx_tokens":484030,"idle_secs":92500}
The session's task text is deliberately not recorded. It is free-form prose out of a transcript, and an audit trail of what was stopped should not become a copy of what was being worked on.
Because each record carries the context that session was holding, the log
answers afterwards what a sweep actually reclaimed rather than just how many
rows you closed. --no-log
records nothing. A log that cannot be written is ignored rather than raised — losing the log is survivable, losing the display is not.
Subagents have no process of their own — they run as sidechains inside the parent's process. Every pid-based view is structurally blind to them.
They do each get a transcript, one directory deeper than the session transcripts:
~/.claude/projects/<slug>/<sessionId>/subagents/agent-<id>.jsonl
The short task description ("Scout source URLs") lives only in the parent's
toolUseResult
, keyed by agentId
. roost joins the two, and falls back to the opening words of the subagent's own first message when the parent's record has scrolled out of reach.
Local, read-only sources. Nothing is sent anywhere; there is no network call except a localhost probe of the infra ports.
| source | gives |
|---|---|
~/.claude/sessions/<pid>.json |
|
| Claude Code sessions: pid, sessionId, cwd, name | |
~/.claude/projects/*/<sid>.jsonl |
|
| Claude model in use, token usage | |
~/.cursor/projects/*/agent-transcripts/*/*.jsonl |
|
| Cursor composers: model, usage, task text (no pid) | |
127.0.0.1 ports |
|
| ollama / litellm / openwebui |
Backends default to both Claude Code and Cursor (ROOST_BACKENDS=claude,cursor
).
Set ROOST_BACKENDS=claude
for Claude-only. Cursor rows show in the SRC
column when both are present. Cursor composers are read-only in interactive
mode — x
cannot stop them (there is no pid), and y
copies the composer id.
Cursor transcripts are supplementary on disk; roost prefers the
composerHeaders
table in Cursor's state.vscdb
for name and context %,
and falls back to JSONL tails when the DB is missing. Composers older than
ROOST_CURSOR_MAX_IDLE_SECS
(default 24h) drop off the board. See
docs/cursor-on-disk.md
and epic
#48.
python scripts/cursor_recon.py
inventories the live layout.
Context is the last assistant turn's `input_tokens + cache_read_input_tokens
- cache_creation_input_tokens`. Cross-checked against an independent tool on the same session: 77% vs 77.29%.
The context window is inferred, not recorded. Nothing on disk states which
window a session opened with, and a session on the 1M window will read 480k+
cache tokens in a single call — scoring that against 200k yields a nonsense
"242%". roost picks the smallest standard tier the usage fits and prints it in
the WIN
column, so the assumption is visible rather than silent. If a new tier
ships, WINDOW_TIERS
is the one line to edit.
- It reads an undocumented on-disk format that can change without warning. That is the whole foundation; treat breakage as expected, not exceptional. - The window inference above is a heuristic.
- The
ADVICE
panel's thresholds are tuned to one person's usage. ReadEXPENSIVE_TOKENS
and friends before trusting the advice. - Daily-driven on macOS and Windows. CI runs the suite and a smoke frame on Linux, and it detects live sessions there — but nobody lives on it yet. Field reports welcome.
roost is one person's tool with a public issue tracker. The help wanted issues are real asks, not decoration: Linux field reports, a canary test for the undocumented on-disk format, adapters for other agent CLIs (gated on demand — comment if you would use one), and config seams for the window tiers and the ADVICE thresholds.
Questions or ideas? → Discussions
On COOPER the working copy lives at dev/roost
(alongside dev/heron
,
dev/swamp-ops
, and the rest of the estate). GitHub still hosts the remote;
clone or copy there if you prefer GitHub/roost
.
roost.py
— the entire program. One file, stdlib only.bin/roost.js
— the npm wrapper; locates a Python interpreter and runsroost.py
under it.packaging/
— the Homebrew formula (roost.rb
), the.deb
build (build-deb.sh
), the version-consistency check that CI runs on every PR (check-version-consistency.sh
), and the apt signing key (apt/
).tests/
— the unittest suiteci.yml
runs on Linux, macOS and Windows.scripts/
— maintenance utilities (cursor_recon.py
inventories Cursor's on-disk layout for adapter work).demo/
— the vhs tapes and GIFs in this README, plus the fleet stager that produces the staged data they record against..github/workflows/
—ci.yml
(tests, packaging checks, semgrep),release.yml
(publishes a tagged release to PyPI, npm and the Homebrew tap),release-please.yml
(maintains the release PR), andpr-title-lint.yml
(enforces conventional commit titles on PRs).
See CHANGELOG.md.
MIT — see LICENSE. Contact: dev@swamplink.com
Built in a Digital Swamp. From my swamp to yours.