atch is a small C utility that lets you attach and detach terminal sessions,
similar to the detach feature of screen/tmux — but without the terminal emulation,
multiple windows, or other overhead.
The key property of atch is transparency. It does not interpose a
terminal emulator between you and your program. The raw byte stream flows
directly from the pty to your terminal, exactly as if you had run the program
in a plain shell. This means:
- Mouse works. Mouse reporting, click events, and drag sequences pass
through unmodified. No escape-sequence re-encoding, no
set -g mouse on, no fighting with terminfo databases. - Scroll works. Programs that use the alternate screen buffer, or that
emit their own scroll sequences, behave identically inside and outside an
atchsession. There is nothing to configure. - Colors and graphics work. True-color, sixel, kitty graphics, OSC sequences — all pass through untouched.
$TERMis unchanged.atchdoes not set or override your terminal type. The program sees exactly the same$TERMyour shell uses.
In contrast, tmux and screen implement their own terminal emulators.
They re-encode the output stream, which frequently breaks mouse support,
scroll behavior, and newer terminal features unless you find and apply the
right obscure configuration knob — and then remember it on every new machine.
With atch there is nothing to remember, because there is nothing in the way.
When a program runs inside an atch session it is protected from the
controlling terminal. You can detach from it, disconnect, and re-attach later
from the same or a different terminal, and the program keeps running
undisturbed.
Session history survives everything. Every byte written to the terminal
is appended to a persistent log file on disk. When you re-attach — whether
the session is still running, crashed, or you have rebooted the machine — the
full output history is replayed to your terminal first, so you can see exactly
what happened and pick up right where you left off. No plugins, no
configuration, no manual script wrappers. Other session managers keep
history only in memory: when the process dies or the machine reboots, the
output is gone. With atch it is on disk until you clear it.
- Attach and detach from running programs
- Multiple clients can attach to the same session simultaneously
- No terminal emulation — raw output stream is passed through unchanged
- Sessions persist across disconnects, crashes, and reboots
- Full session history on disk — every line ever written is saved and replayed on re-attach
- History survives process exit — re-opening a session shows the complete prior output before starting fresh
- Push stdin directly to a running session
- List sessions with liveness status;
list -aalso shows exited sessions that still have a log on disk - One-command cleanup —
rm <session>orrm -aremoves stale/exited sessions and their logs - Share a session with other users — read-only by default; grant input per-user or per-group, time-boxed and revocable. Access is decided by kernel-verified peer identity (
SO_PEERCRED), not file permissions - Remote sessions with nothing to preinstall —
atch remote host nameself-deploys the static binary over your own ssh, then re-attaches by bare name. No credentials are stored - Prevents accidental recursive self-attach
- Tiny and auditable
The easiest way to get atch is to download a pre-built binary from the
GitHub releases page. The release
binaries are statically linked against musl libc — they run on any Linux
distribution with no dependencies. Just download and extract:
arch=$(uname -m | sed 's/x86_64/amd64/;s/aarch64/arm64/')
curl -Lo atch.tgz https://github.com/mobydeck/atch/releases/latest/download/atch-linux-${arch}.tgz
tar -xzf atch.tgz atch
sudo mv atch /usr/local/bin/
rm -f atch.tgzOr download the .tgz from the releases page and extract it manually.
atch sits between your keyboard and your running program at the lowest
level of the Unix process model: it opens pseudo-terminals, manages Unix
domain sockets, and forwards raw byte streams. These are system-level
operations that map directly to C system calls. There is no meaningful
runtime overhead to hide, no framework to bring in, and no interpreter to
start — the binary does exactly what it advertises and nothing else.
The release binaries are statically linked against
musl libc. A static musl binary is a
single self-contained file that carries everything it needs. You copy
it to any Linux machine — Ubuntu, Debian, Alpine, RHEL, a bare container
image, a two-year-old distro — and it runs. There are no shared library
version mismatches, no apt install, no LD_LIBRARY_PATH gymnastics. For
a small tool you want to drop onto servers and forget about, a static binary
is the right default.
makeatch [<session> [command...]] Attach to session or create it (default)
atch remote <host> <session> Attach to a session on a remote host (bootstrapping it)
atch <command> [options] ...
Sessions are identified by name. A bare name (no /) is stored as a socket
under ~/.cache/atch/. A name containing / is used as-is as a filesystem path.
A bare name that has been registered against a remote host (see
Remote sessions) resolves to that host instead.
If no command is given, $SHELL is used.
| Command | Description |
|---|---|
atch [<session> [cmd...]] |
Attach to a session, or create it if it doesn't exist (default behavior). Prints a confirmation when a new session is created. |
attach <session> |
Strict attach — fail if the session does not exist. |
new <session> [cmd...] |
Create a new session and attach to it. Prints a confirmation before attaching. |
start <session> [cmd...] |
Create a new session, detached (atch exits immediately). Prints a confirmation on success. |
run <session> [cmd...] |
Like start, but atch stays in the foreground instead of daemonizing. |
push <session> |
Copy stdin verbatim to the session. |
kill [-f] <session> |
Gracefully stop a session (SIGTERM, then SIGKILL after 5 s if needed). With -f / --force, skip the grace period and send SIGKILL immediately. |
clear [<session>] |
Truncate the on-disk session log. Defaults to the current session when run inside one. |
tail [-f] [-n N] <session> |
Print the last N lines of the session log (default: 10). Works for both running and exited sessions. With -f, follow new output as it is written (useful for monitoring a running session without attaching). |
list [-a] |
List sessions. Shows [attached] when a client is connected, [stale] for leftover sockets with no running master. With -a, also shows [exited] sessions that have a log file but are no longer running. Prints (no sessions) when the list is empty. |
rm [-a] [<session>] |
Remove a session and its log. Refuses if the session is currently running — use kill first. Works on stale (socket exists but process is dead) and exited (log only) sessions. With -a, sweeps all stale and exited sessions at once. |
share <session> --to <spec> [-m MIN] [--write] |
Share a running session with other local users, read-only by default. <spec> is a comma-separated list of user or @group targets, each with an optional :rw / :ro suffix; --write makes every target writable by default. -m auto-revokes after MIN minutes (0 = never; default 60). See Sharing a session. |
join <session> |
Attach to a session another user has shared with you. |
unshare <session> |
Revoke a share immediately. Expiry revokes it automatically as well. |
remote <host> <session> |
Bootstrap (on first use) and attach to a session on a remote host over ssh. <host> may be a hostname, FQDN, IP address, or user@host. See Remote sessions. |
remote <add|ls|rm> ... |
Manage the remote-session registry: add <name> <host> maps a name to a host, ls lists the mappings, rm <name> forgets one. |
current |
Print the current session name and exit 0 if inside a session; exit 1 silently if not. |
Short aliases: a → attach, n → new, s → start, p → push,
k → kill, l / ls → list, j → join.
Options can appear before the subcommand, before the session name, or after the session name.
| Flag | Description |
|---|---|
-e <char> |
Set the detach character. Accepts ^X notation. Default: ^\. |
-E |
Disable the detach character entirely. |
-r <method> |
Redraw method on attach: none, ctrl_l, or winch (default). |
-R <method> |
Clear method on attach: none or move. |
-z |
Disable suspend-key (^Z) processing (pass it to the program instead). |
-q |
Suppress informational messages. |
-t |
Disable VT100 assumptions. |
-C <size> |
Set the on-disk log cap for the session being created. Accepts a bare number (bytes), or a number with k/K (KiB) or m/M (MiB) suffix. 0 disables the log entirely. Default: 1m. |
Use -- to separate atch options from command arguments that start with -:
atch new mysession -- grep -r foo /var/logStart a shell session named work and attach to it:
atch workStart a specific command in a named session:
atch new build -- make -j4Attach to an existing session, creating it if needed:
atch workStrict attach — fail if the session is not running:
atch attach workDetach from a running session: press ^\ (Ctrl-\). The session and its
program keep running.
Re-attach later:
atch workRun a command fully detached (no terminal needed):
atch start daemon myserver
# atch: session 'daemon' startedUse -q to suppress confirmation messages in scripts:
atch start -q daemon myserverSend keystrokes to a running session:
printf 'ls -la\n' | atch push workUse a custom detach character:
atch -e '^A' attach workList all sessions:
atch listList sessions including those that have exited but still have a log:
atch list -aInspect the last 20 lines of a session log:
atch tail -n 20 workFollow a session's output without attaching:
atch tail -f workKill a session:
atch kill workRemove a stale or exited session (socket + log):
atch rm crashed-jobSweep all stale and exited sessions in one go:
atch rm -aShare a session read-only with a coworker for 30 minutes:
atch share work --to alice -m 30
# alice, on the same machine:
atch join workShare with a whole group, and let one person type:
atch share deploy --to @ops,alice:rwRevoke a share early:
atch unshare workOpen a session on a remote host that has never seen atch:
atch remote 10.0.0.7 logs
atch remote ansible.example.net logs # hostname / FQDN works tooRe-attach to it later by bare name (no host needed):
atch logsBy default, session sockets are stored in ~/.cache/atch/. The directory is
created automatically, including ~/.cache if it does not yet exist.
When $HOME is unset or empty, atch looks up the home directory from the
system user database (/etc/passwd). If that also yields nothing useful (or
points to /), sockets fall back to /tmp/.atch-<uid>/.
To use a custom path, include a / in the session name:
atch new /tmp/mysessionatch sets the ATCH_SESSION environment variable inside each session to a
colon-separated ancestry chain of socket paths, outermost first. A
non-nested session has a single path; nested sessions accumulate:
outer session: ATCH_SESSION=/home/user/.cache/atch/outer
inner session: ATCH_SESSION=/home/user/.cache/atch/outer:/home/user/.cache/atch/inner
This serves two purposes: self-attach prevention (any ancestor in the chain
is rejected) and session detection from scripts. Use atch current to get
the human-readable session name — it prints just the basenames separated by
>:
# exit code: 0 inside a session, 1 outside
atch current && echo "inside session: $(atch current)"
# nested session example:
# outer > inner
# shell prompt example (bash/zsh PS1)
PS1='$(atch current 2>/dev/null && echo "[$(atch current)] ")$ 'To test whether you are inside any atch session:
[ -n "$ATCH_SESSION" ] && echo "inside a session"atch keeps two complementary history stores, both replayed automatically
whenever you attach — no configuration required.
Every byte written to the pty is appended to a log file on disk
(~/.cache/atch/<session>.log). The log persists across everything:
- Detach / re-attach — re-attaching to a running session replays the complete history before the live stream begins.
- Session exit — once the program exits, the full output remains on disk.
Running
atch mysessionagain starts a fresh session but first shows everything from the previous one, so you know exactly what it did. Useatch list -ato see all sessions that still have a log, including exited ones. - Machine reboot — the log file survives a reboot. The next time you open the session you see the complete prior output before the new shell starts.
- Crash recovery — if the session process is killed unexpectedly, the log is intact. Nothing is lost.
This is fundamentally different from tmux, screen, and dtach: they hold
history only in memory. When the process exits or the machine restarts, the
output is gone. With atch the raw byte stream is on disk until you
explicitly clear it with atch clear (or atch clear <session>).
To inspect the log without attaching to the session, use atch tail:
atch tail mysession # last 10 lines
atch tail -n 50 mysession # last 50 lines
atch tail -f mysession # follow live outputThis works whether the session is running, exited, or from a previous boot.
The log is capped at 1 MB by default; once it exceeds that, only the most
recent 1 MB is kept. You can change the cap per session with -C:
atch -C 4m new mysession # 4 MB cap
atch -C 128k start daemon # 128 KB cap
atch -C 0 start daemon # no log at allWhen the log is disabled with -C 0, re-attaching to a running session
still replays recent output from the in-memory ring buffer. Only cold replay
of a dead session (after the master has exited) is unavailable.
To change the compiled-in default, build with:
make CFLAGS="-DLOG_MAX_SIZE=$((4*1024*1024))"While the session is running, atch maintains a 128 KB ring buffer in the
master process. It is the primary replay source when you re-attach: the ring
replays the most recent output instantly so your display is current. When the
on-disk log is also present it covers the full history; when logging is
disabled (-C 0) the ring is the only replay source available while the
session is live.
The ring is lost when the master exits; the on-disk log covers that case.
To adjust the ring size, build with:
make CFLAGS="-DSCROLLBACK_SIZE=$((256*1024))"The value must be a power of two.
To wipe the on-disk log and start clean on the next attach:
atch clear # inside a session — clears the current session's log
atch clear mysession # from outside — clear a named session's logTo completely remove a session that is no longer running (deletes both the socket file and the log):
atch rm mysession # remove one stale or exited session
atch rm -a # remove all stale and exited sessions at oncerm refuses to touch a running session — use atch kill first if you want to stop and remove it.
To clear the log automatically whenever you run clear, add a shell function
to your .bashrc / .zshrc:
clear() { command clear; [ -n "$ATCH_SESSION" ] && atch clear 2>/dev/null; }This only fires for the literal clear command, not for full-screen programs
like vim or htop that also erase the terminal.
By default a session is yours alone: its socket is chmod 0600 in your own
~/.cache/atch/, so the filesystem already keeps everyone else out. atch share opens a running session to specific other users on the same machine —
read-only by default, so the common case of "let someone watch what I'm
doing" carries no risk of them touching the keyboard.
atch share work --to alice # alice can watch, not type
atch share work --to alice:rw # alice can type too
atch share work --to @ops # everyone in group "ops", read-only
atch share work --to alice:rw,@ops # mix per-target modes freely
atch share work --to @ops --write # writable by default for every targetA share is time-boxed. It auto-revokes after 60 minutes by default; pass
-m <minutes> to change it, or -m 0 for no expiry. atch unshare work
revokes it immediately. Either way the guest listener is closed, its socket
removed, and any connected guests are dropped on the spot.
A coworker the session is shared with attaches with atch join:
atch join workMultiple guests can join at once, and they see the full replayed history just like the owner. Read-only guests have their keystrokes silently discarded — they cannot inject input or signal the program no matter what they send.
Guests don't need atch installed. This matters most for a session you are
running on a remote host (atch remote …): the bootstrapped binary lives in
the session owner's ~/.cache/atch/, which another user's account cannot read.
So when you share, atch also drops a world-readable copy of itself that any
local user can exec by absolute path, and share prints the exact command:
atch: session 'work' shared to 1 target(s), expires in 30m 0s
guest socket: /tmp/.atch-guest.1000.work
others can run: atch join work
(no atch on their box? /tmp/.atch-bin.1000 join work)
A root-owned session stages it at /run/<prog>/<prog> (e.g. /run/atch/atch);
a non-root session uses /tmp/.<prog>-bin.<uid>, because creating a directory
under /run needs privilege and atch stays unprivileged. So the guest simply
runs the printed command, e.g. /run/atch/atch join work.
Authorization is by identity, not by file permissions. When you share a
session, atch spins up a second guest listener socket in a world-traversable
directory so other users can reach it; that socket is deliberately left
wide-open at the filesystem level. The real gate is in the master: on every
connection it reads the peer's user and group identity straight from the kernel
(SO_PEERCRED) — an identity the connecting process cannot forge — and checks
it against the targets you granted, including the peer's supplementary groups.
Anyone not on the list is refused before they see a single byte. The grant, and
every join and rejection, is recorded in the session log. When the share ends
the listener disappears, so there is never an open socket sitting around without
an allow-list behind it.
This needs no privileged setup — no sudo, no shared group you both have to
belong to, no fiddling with socket ownership. atch never calls setuid and
runs entirely as you.
atch can attach to a session on another machine that has nothing
preinstalled — not even atch itself:
atch remote 10.0.0.7 logs # IP address
atch remote ansible.example.net logs # hostname or FQDN
atch remote deploy@10.0.0.7 logs # or user@hostOn first contact atch bootstraps the host over your own ssh — whatever
already lets you ssh 10.0.0.7 (an agent key, an entry in ~/.ssh/config, or a
password prompt) is exactly what it uses. atch stores no credentials of its
own. Over that one connection it:
- generates a dedicated ed25519 key (
~/.config/atch/id_ed25519) the first time it is ever needed, and installs it into the remote~/.ssh/authorized_keysbehind a restricted, forced-command line so the key can do nothing but invoke the relay; - copies the self-contained static binary into the remote
~/.cache/atch/, so the host needs no package, no compiler, and nothing in itsPATH; - pins the host's key on first sight (trust-on-first-use) in a private
~/.config/atch/known_hosts, and refuses to continue if it ever changes; - records the
name → hostmapping in~/.config/atch/registry.
After that, the bare name is enough — no host, no re-bootstrapping, and no interactive auth, because the dedicated key takes over:
atch logs # re-attaches to 'logs' on 10.0.0.7The registry holds no secrets, only the host, its pinned fingerprint, and the
path to the identity key. Manage it directly with atch remote:
atch remote ls # list registered remote sessions
atch remote add logs 10.0.0.7 # record a name → host mapping by hand
atch remote rm logs # forget a mappingUsing a dedicated key rather than your personal identity, locked to a forced command, keeps the blast radius to "can run atch on this host" if it ever leaks.
Architecture must match; kernel and distro don't. atch remote ships your
local binary as-is — it does not recompile on the far side. Because the binary
is fully static, it runs on essentially any Linux kernel and distribution. The
one thing that must line up is the CPU architecture. atch detects the remote's
(uname -m) and, when it differs from yours (say an arm64 box from an x86_64
workstation), stages a matching static binary from ~/.config/atch/<prog>-<arch>
(amd64 / arm64) instead of itself; build the per-arch artifacts with make release and drop the right one there. If none is present it refuses up front
with a clear message rather than shipping an unrunnable binary.
Once you are attached, atch is on your PATH inside the session even on the
remote host — the master prepends its own directory (e.g. ~/.cache/atch/) to
PATH for the session — so you can run atch share, atch list, and the other
subcommands directly from the remote shell without spelling out the full path to
the binary.
The original flag-based syntax is still supported:
atch -a <session> # same as: atch attach <session>
atch -A <session> [cmd...] # same as: atch [<session> [cmd...]]
atch -c <session> [cmd...] # same as: atch new <session> [cmd...]
atch -n <session> [cmd...] # same as: atch start <session> [cmd...]
atch -N <session> [cmd...] # same as: atch run <session> [cmd...]
atch -p <session> # same as: atch push <session>
atch -k <session> # same as: atch kill <session>
atch -l # same as: atch list
atch -i # same as: atch current
Existing scripts do not need to be updated.
GPL. Based on dtach by Ned T. Crigler.