Quickstart

1. curl -fsSL https://raw.githubusercontent.com/mmr710/nightmux/main/install.sh | sh
2. Paste a bot token from @BotFather — or press Enter to run dashboard-only.
3. In a topic (or + project on the dashboard): !new myproj ~/code/myproj, then just type.

Just looking? python3 nightmux.py --demo and open http://127.0.0.1:8099/office — no bot, no keys. Docker, a fresh VPS and the rest: Install →

How it compares

	nightmux	hosted / relay mobile apps	tmux + SSH from a phone
Resumes by itself after a usage limit, or hands off to another agent	✅	—	—
Any agent CLI (Claude Code, Codex, Gemini, opencode, Cursor…)	✅	usually one	✅
Phone UI: menus as buttons, photos, voice notes	✅	✅	—
Runs on your machine; no relay sees your code	✅	—	✅
Overnight: schedules, !goal test loops, morning briefing	✅	—	—
Attaches to the tmux session you can still SSH into	✅	—	✅
One file, Python stdlib, no account	✅	—	✅

The night shift

02:14  ⏸ api hit the usage limit
       5-hour window spent — resumes 04:11, resuming itself with 'continue'
04:11  ▶️ api resumed · sending queued prompt
04:11  ⚙️ api

A usage limit at 2am used to end the night. The turn dies mid-refactor, the prompt that started it is already spent, and the session sits there until someone awake types continue. nightmux reads the reset time, holds everything you send, and puts the work back the moment the window reopens — including the turn the limit cut off. You read the result at breakfast.

That is the part nobody else is doing. The rest is what makes it usable:

Run Claude Code from your phone — or Codex, Gemini, aider, anything with a prompt. One Telegram forum topic per project, one tmux session behind it. Text you send is typed into that session's prompt; what the session says comes back to the topic. Approvals arrive as tap buttons.

No container, no DNS, no certificates, no ports open, no relay service. It attaches to tmux sessions you already have, on the machine you already use. Python stdlib only — one file you can read.

   Telegram group (Topics on)          your machine
   ┌───────────────────────┐          ┌──────────────────────────┐
   │ #api      ────────────┼──────────┼─► tmux: api    → claude  │
   │ #frontend ────────────┼──────────┼─► tmux: web    → claude  │
   │ #scratch  ────────────┼──────────┼─► tmux: scratch→ claude  │
   └───────────────────────┘          └──────────────────────────┘
              ▲                                    │
              └──── output, approvals, usage ──────┘

Why this one

There are plenty of ways to reach a coding agent from a phone. Most are one of two shapes: a bot that drives the agent through its SDK and keeps the conversation in its own database, or a mobile app that talks to a relay service you don't run. Both work. Neither leaves you with a terminal session.

nightmux is the third shape — it drives the session you would have started yourself:

It works the hours you don't. A status-line sidecar gives nightmux the real context percentage and the real 5-hour / 7-day limit windows, so it can act on them instead of discovering them:

• a turn the limit cut off resumes itself when the window reopens ("auto_continue": false to wait for a human instead)
• a prompt sent during a lockout is held, not lost — replayed when the window resets, surviving daemon restarts and reboots
• a prompt refused before it ever got a turn goes back on the queue whole
• !at 03:00 <prompt> and !every 4h <prompt> start work while you are asleep — and they queue rather than type, so they wait behind a lockout too
• /compact automatically at a context threshold you set (!autocompact 70)
• warnings at 80% and 90% of a window, before the wall rather than at it
• !ctx shows what is actually filling the window; !cost weighs a session or every project by token type

Long-running agent sessions cost money and stall in ways chat never does. That is the part nobody else is watching.

It attaches to sessions instead of owning them. nightmux types into tmux. The session is still yours — SSH in, attach, type directly, and the bot keeps working mid-conversation. Nothing is wrapped, proxied, or re-hosted, so there is no state to get out of sync and nothing to lose when the daemon restarts.

It is not tied to one agent. !new starts your default; !codex, !agy, !opencode, !gemini, !aider, !cursor (Cursor CLI), !amp, !goose, !qwen (Qwen Code) or anything you add to agents in the config starts that instead, and !resume remembers which agent a topic belongs to. The hooks and the usage numbers are Claude Code specific — every other agent degrades to reading the terminal, which is how nightmux worked before the hooks existed.

One topic, several agents. A bare !agy, !codex or !opencode in a topic that already has a directory switches that topic to that agent — same project, its own tmux session, and the agent you were on left running. !agents lists the bench and marks the live one; switching back lands in the conversation it was in, not a fresh one. Give a name and a directory (!agy side ~/code/api) and it still means start-a-new-session, as before.

One session belongs to one topic. Two topics pointing at the same session share a single scrape cursor, so whichever one the watcher reaches first gets the output and the other goes quiet; !bind refuses that now, and !status flags any pair already in your config.

Not for you if you want a polished app instead of a chat window, you're on Windows (#2 — macOS and Linux both install as a service), or you want your teammates in the same group: the allowlist is a list of people trusted with a shell on your machine, which is not a thing to hand out. One person, their own box, their own agents.

Install

The fastest way to install is using the one-line installer:

curl -fsSL https://raw.githubusercontent.com/mmr710/nightmux/main/install.sh | sh

(Clones — or updates — ~/nightmux from GitHub and runs setup. Adding a second server? Use the command its dashboard gives you instead: servers → + add server.)

Just looking? python3 nightmux.py --demo serves the office with a made-up night — no bot, no keys.

Docker (tmux, Claude Code and Codex inside; your code mounted at /code):

docker run -it --name nightmux -v nightmux-home:/root -v ~/code:/code \
  -p 127.0.0.1:9090:9090 ghcr.io/mmr710/nightmux
docker exec -it nightmux claude      # log the agent in once

The dashboard has no login of its own — keep the port on 127.0.0.1 (or your tailnet).

A fresh VPS (Hetzner, DigitalOcean, Vultr…): paste deploy/cloud-init.yaml into the provider's user-data box, then SSH in, log the agents in and run the one-liner above.

Or from PyPI (may lag behind GitHub — the peer and dashboard features need the current version):

pipx install nightmux
nightmux --setup

Or install the latest development version directly from GitHub:

pipx install git+https://github.com/mmr710/nightmux
nightmux --setup

or clone it, which is the version to pick if you want the source where you can read and edit it — there are only four files and no dependencies:

git clone https://github.com/mmr710/nightmux ~/nightmux
python3 ~/nightmux/nightmux.py --setup

Setup walks the whole thing: BotFather token, finding your group, writing the allowlist, wiring the Claude Code hooks, installing the service (a systemd user unit on Linux, a launchd agent on macOS). It is idempotent — re-run it after an upgrade.

You will be asked to create a Telegram group with Topics turned on and add the bot as an admin. Admin is not optional: without it the bot only receives messages addressed to it, so most of what you type never arrives.

Then, in a new topic:

!new api ~/code/api      # start a session and bind this topic to it

and type. !help lists the rest.

What it feels like

you   fix the failing auth test
bot   ⚙️ api  · Opus 5 · 34% ctx
bot   🔧 Bash  pytest tests/test_auth.py -x
bot   🔧 Read  src/auth.py
bot   🟠 needs input api
      Bash(git commit -m "fix token expiry check")
      [ 1. Yes ] [ 2. Yes, don't ask again ] [ 3. No ]
you   (taps 1)
bot   ✅ api
      Token expiry used `<` instead of `<=`, so a token expiring exactly on the
      boundary was rejected. Fixed and committed; the test passes.

Approvals arrive the moment Claude Code asks, via its Notification hook — before the terminal has finished redrawing.

Commands

Everything works as !cmd, and the common ones are registered as /cmd so Telegram autocompletes them. Anything that is not a nightmux command — including Claude's own /compact, /clear, /model — is typed into the session.

!new <name> [dir] [flags] [@branch]	start a session with the default agent, bind this topic to it
!codex / !aider / !gemini / !agy …	same, with that agent
!<agent> with no arguments	switch this topic to that agent, in the same directory
!agents	this topic's agents, which one is live, and which report a context figure
@claude <text> / @agy <text>	send one prompt to one agent on the bench, without switching the topic
!consult <question>	ask every agent separately, let them read each other, get one prompt back
!use [agent]	run the prompt a consult settled on, in this topic's agent
!resume [agent] / !restore	relaunch this topic's directory, resuming the last conversation
!bind <session> / !unbind / !kill	attach, detach, stop (kill asks first)
!sessions / !status	tmux sessions; every topic and its state
!pane [lines] / !ctl	dump the terminal; button panel
!git / !diff / !get <path>	repo state and file upload from the session's cwd
!worktrees	git worktrees of this topic's repo, and which session sits in each
!undo	list snapshot branches for this repo and the restore commands — never runs them
!ctx / !cost [days] / !usage	context breakdown, token spend, limit windows
!autocompact <pct|off>	auto-/compact at a context threshold
!spendcap <turns|500k|2M|off>	interrupt a session that runs up that many turns, or that many base-equivalent tokens, in 5 minutes
!idlectx <pct|off>	flag parked sessions still holding a big context
!queue [clear|now]	prompts held for a rate-limit reset
!at 03:00 <prompt>, !at +90m …	run a prompt later
!every 4h <prompt>, !sched [clear]	run it on a repeat, or list what is set
!shift + one prompt per line	a sequential overnight plan, one turn at a time
!digest [HH:MM|off]	what happened while you slept, on demand or daily
!center [off]	make this topic watch and control every session
!board	every topic at a glance, from any topic
!all <targets|--all> <prompt>	send one prompt into several sessions at once
!grep <text> [days]	search every transcript on the machine
!autoyes <agent|off>	answer one agent's own permission menus for it, per topic, off by default
!verbose / !raw <text> / !keys <keys>	tool detail, type past a menu, raw tmux keys
!1..!9 !y !n !esc !int !enter !tab	menu picks and keys

A pick is checked, not assumed: !1 sends the digit, looks at the pane, and adds Enter only if the same question is still there — dialogs disagree about whether a digit confirms or only moves the highlight. !y/!n answer a numbered menu with the digit of its Yes/No option, because the letter does nothing to a list. | !version | build, python, and which hooks are wired | | !tz <zone> / !reload / !log / !help | timezone, re-read config, journal, this list |

Send a photo, file or voice message and it is saved, with the path typed into the session.

Several agents, one project

!new api ~/code/api @refactor-auth checks out a git worktree for that branch — new if it doesn't exist, reused if it does — at ~/code/api-wt/refactor-auth, and starts the session there instead of in the main tree. Two agents, two branches, no stepping on each other's uncommitted work. !worktrees lists every worktree of the current topic's repo and which session (if any) is sitting in it. Nothing here routes work between agents or merges anything — that part is still yours; this is just isolation.

Consulting two agents at once

!consult <question> puts the same question to every agent on the topic's bench, separately. Round one goes out blind — showing each of them the other's answer first gets agreement, and agreement between two models that read each other is worth much less than two independent reads. Round two hands each one what the other said and asks for a single self-contained prompt, fenced.

What comes back is one message per prompt, or one message if they converged on the same wording. The drafts in between are not posted: they are working notes, and both of them arrive in the report anyway.

!consult should the parser stream or buffer?
🤝 consult round 1 — asking agy, claude, separately
🤝 consult round 2 — each now reads the other
🤝 consult done — agy, claude
   they differ; both are below, yours to pick

What comes back carries a button. !use — or a tap on run it — sends that prompt into the agent this topic is on, so the answer to "what should I ask for" becomes the thing being asked without anyone retyping it on a phone. When the two disagree you get one button each, and !use agy / !use claude pick between them.

A consultation lives in memory only. Restarting the daemon cancels it, on purpose: half a conversation restored into two sessions that have since moved on is worse than none.

Command center

!center in any topic makes it the one place that watches and controls every session — it binds to nothing itself, !board there shows every topic's state at a glance:

you   !board
bot   ✅ api      topic 12   idle    2s quiet  5h 40%  ctx 22%
      ⚙️  web      topic 15   busy    0s quiet  5h 40%
      🟠 scratch  topic 19   waiting 1s quiet  🔒held→04:11
      spend  api ~12,400tok · web ~3,100tok

Approvals mirror there too: a session's 🟠 needs-input prompt posts to its own topic as always, and a copy lands in the command center with the same buttons. Whichever is tapped first answers the pane; the other loses its buttons immediately rather than sitting there able to send a second, conflicting keystroke. !all web,scratch --continue (or !all --all <prompt> for every bound, writable session) sends the same prompt to several sessions at once — it replies with exactly who it is about to hit before anything is typed, and a "readonly" topic is never one of them. Nothing here routes work between agents or decides anything for them; it broadcasts and it aggregates, and every session still runs on its own.

Snapshots and !undo

Before nightmux sends a prompt it typed without you watching it happen — a prompt held for a usage-limit reset, an !at/!every firing, a !shift step — it takes a git snapshot of the session's cwd first: git stash create captures the worktree without touching it, and a nightmux/pre-<UTC timestamp> branch is left pointing at it (the last 5 per repo; older ones are dropped). Prompts you type live get no snapshot — there is nothing unattended about them. !undo lists a topic's snapshot branches, newest first, with the exact git restore/git diff commands to look at or roll back to one. It never runs them — a phone is a small thing to fat-finger a hard reset from.

Zero-Dependency Webhook API + Dashboard

nightmux runs a local HTTP server (127.0.0.1:9090) to accept commands from outside Telegram. You can configure "webhook_port": 9090 in your ~/.nightmux.json to enable it.

This turns nightmux into the central nervous system for your local agents. You can pipe GitHub Actions test failures or VS Code compiler errors straight into your agent's queue while you sleep.

curl -X POST http://127.0.0.1:9090/topic/api -d "review the staged changes"

The same port also serves a page — open http://127.0.0.1:9090/ for every bound topic at a glance (mode, usage, who else is on the bench) with a box to send a prompt, no phone required. /api/topics is the JSON it polls, if you want to build your own view instead.

Check out the Cookbook for copy-paste recipes for GitHub Actions and editor integrations.

From idea to running project

In a new topic: !idea a habit tracker with streaks and a weekly chart (!idea @codex … to pick the agent). nightmux makes a fresh folder under projects_root, runs git init, starts the agent with one queued brief — write SPEC.md, build the smallest useful version, add a ./check.sh that runs the tests — and turns on !goal sh ./check.sh, so it keeps going until its own check passes. !preview when it serves something, !p ship when you like it.

Saved prompts: !p shows review, fix-tests, spec, explain, tidy and ship as buttons; !p save <name> <text> adds yours ({{args}} marks where !p <name> <args> puts the rest).

Previews and production errors

Deploy previews: while !watch follows a PR, the preview URL your host attaches to its commit — a GitHub deployment (Vercel, Render…), a deploy-preview status (Netlify) or a link in a bot's comment (Cloudflare Pages, Netlify) — is posted once per commit with 🔗 open and 📸 screenshot (!shot last); a photo after that is a point-and-fix report about the preview. Bot comments are no longer sent to the agent as review feedback.

Errors: !errors in a topic gives it a webhook, https://<host>:8443/hook/<topic>?key=…. Point a Sentry custom (internal) integration's alert action at it — or POST any JSON {title, stack, url} — and each new error (deduplicated for 6 hours) arrives as 🚨 with a fix it button; !errors auto sends it straight to the agent: reproduce with a failing test, fix, push. Sentry is on the internet, so !errors expose (you run it) publishes only /hook, on port 8443, with Tailscale Funnel — the dashboard stays tailnet-only and the hook refuses anything without the key. !errors unexpose takes it back off the internet.

Issues in, pull requests out

!issues lists the repo's open issues as buttons. Tap one (or !issue 12): the agent gets the issue and its comments, a branch name (issue-12-…) and the recipe — branch, fix with a test that fails first, push, gh pr create ending in Fixes #12. nightmux watches for that branch's PR and, once it opens, !watch takes over: CI failures and review comments go back to the agent, green gets a merge button. !issues auto (label nightmux, or !issues auto <label>) works through labelled issues one at a time whenever the topic has been quiet 10 minutes with nothing in flight — label a few before bed. !issues auto off stops.

Close the PR loop

!watch pr (this branch's PR) or !watch pr 12: nightmux polls it with gh. A failed check's log goes to the agent — fix, commit, push. New review comments, inline ones included, go to the agent too; comments made by your own gh account are skipped so an agent replying on the PR never feeds itself. When every check is green you get one 🟢 with a merge button. Merged or closed ends the watch.

Android, on your phone

!apk builds the project's debug APK — a Gradle wrapper at the root or under android/ (React Native, Capacitor), or Flutter — and sends it to the topic: tap to install. With your phone on wireless debugging (Developer options → Wireless debugging, reachable over the tailnet or Wi-Fi): !android pair <ip:port> <code> once, !android connect <ip:port>, then !apk install builds and installs in one go, and !shot android sends the phone's screen. A photo sent after that is a bug report: the agent gets it with the activity on screen and recent logcat errors. A failed build has a send to agent button. nightmux uses the first adb on PATH that actually runs (an SDK's x86 adb on an ARM server does not; apt install adb).

A desktop for agents

!desktop starts a virtual screen on the server (Xvfb + openbox), and gives you a 🖥 button: noVNC over your tailnet, behind a VNC password, to watch or take over from the phone. !tools add desktop gives the topic's agent nightmux's own MCP server for it — screenshot, click, type, key, scroll, open — so it can drive GUI apps, a real browser with a login, an emulator. !desktop open <app> starts something yourself, !desktop shot sends a screenshot, !desktop off stops it all. Needs xvfb openbox x11vnc novnc websockify xdotool imagemagick (apt).

Give the agent a browser

!tools add browser registers Playwright's MCP server with the topic's live agent through its own CLI (claude mcp add -s local — this project only — or codex / agy / opencode mcp add); !tools add browser all does every agent on the bench. Tap restart (or !tools restart) so the agent loads it, and it is told to use it: open the running app, click through what it changed, read the console, before it says done. Headless, sandbox-off for servers, and it reuses the Chromium nightmux already found (Playwright's, or "browser" in the config). !tools rm browser removes it.

See it on your phone

!preview finds the dev server the agent started (from the session's own process tree) and replies with a 📱 button per app. A server on all interfaces is linked on your tailnet directly; one on 127.0.0.1 gets a tailscale serve on the same port. Nothing is exposed to the internet. !shot sends a phone-size screenshot of it (!shot :5173/settings, or any URL) — uses Playwright's Chromium if present, else the system's, or "browser" in the config.

Point and fix: within two hours of a !preview or !shot, a photo you send to the topic — a screenshot with the bug circled, captioned or not — is treated as a bug report about that page. The agent gets the image, the page URL, that page's console errors and its rendered DOM (as a file), and is told to find the code behind what is marked, fix it, and load the page again.

The right agent for the task

!route auto in a topic with several agents on its bench: a prompt that starts a new task (the topic has been quiet 10 minutes, no !goal loop running) is sorted light (rename, typo, add tests, docs), heavy (design, debug, why, refactor, performance, security) or normal, and the topic switches to the best live agent for it — ↪️ → codex (light task). Mid-task prompts never move. !goal results teach it which agent actually finishes which kind of task; "route_prefs" in the config sets the starting order. @agent <prompt> still overrides; !route off stops it.

The night, as a GIF

nightmux records the office's desks once a minute (only when something changes). !reel (last 12h; !reel 8 for 8) replays the busiest room as a GIF — drawn by the office page itself in headless Chromium, stitched by a small stdlib GIF encoder — captioned with what happened ("claude hits its limit, codex gets to work") and each project's commits and lines changed. !reel daily 07:30 sends it every morning; !reel off stops that.

A second pair of eyes

!pair codex in a Claude topic (any two agents): after every turn that changed the tree, the reviewer — started on the topic's bench if it is not running, never switched to — runs git diff <since> in the same folder and reports only real problems as file:line — problem — fix, or LGTM. LGTM is a quiet 👍; findings go straight back to the coder, at most 2 rounds per change (!pair codex 4 for more) before it hands the disagreement to you. !pair off ends it.

When an agent goes in circles

Every finished turn leaves a few cheap readings: error lines, which files changed against HEAD, apologies. Two independent signs of circling in the recent turns — the same error 3×, one file edited 4+ turns while the diff does not grow, repeated "I apologize / let me try another approach", the !goal check failing the same way — and you get 🌀 looks stuck with buttons: 🧠 step back (list hypotheses before editing), 🔀 hand to another agent with a summary, ⏸ pause. !loopguard auto lets the first one step it back by itself; !loopguard off disables it.

Morning briefing

!briefing 07:30 in any topic: every morning at that time, one message there — what got done overnight (commits and lines per project), what is waiting for you (agents asking a question, open PRs with GitHub's verdict: ready, behind, conflicting; production errors from !errors), each usage window and when it fills at the current pace, and what is queued, held or stopped. Buttons open the night reel and yesterday's stats. !briefing now for one right away, !briefing off to stop.

Dependency health

!deps runs the project's own tools — npm audit / npm outdated, pip-audit, govulncheck, whichever apply and are installed — and lists known vulnerabilities worst first and packages a major version behind, with one 🛠 upgrade with agent button: fixes first, majors one at a time with their changelogs, tests after each, one commit each (your !goal check, if set, holds it to green). !deps nightly 04:00 checks every night and only speaks when it finds something; !deps nightly off stops it.

See the limit coming

Every usage window an agent reports (Claude's 5h/7d, Codex's) is sampled once a minute. At the last hour's pace, if one will be full before it resets — and within 90 minutes — the topics working on that agent get one ⏳ warning ("claude will hit its 5h limit around 03:10 at this pace, 82% now, +14%/h") with a button to switch to another agent while there is still time. !forecast shows every window; the dashboard's limit cards show the projected time too.

Project memory

Each project keeps .nightmux/memory.md — what it is, how to run and test it, decisions and why, gotchas, what is in progress. After six finished turns and 15 quiet minutes the live agent rewrites it (that exchange is not posted, just a quiet 🧠), and every session nightmux starts or switches to in the topic is told to read it first — so a fresh agent, a failover or a switch does not start from zero, and you stop re-explaining. It is kept out of git through .git/info/exclude. !memory shows it, !memory update refreshes it now, !memory off stops it for the topic.

Spend fewer tokens per task

Every turn re-reads the whole context, so the savings are in turns that never happen and contexts that stay small.

• Auto-compact is on by default at 200k tokens on Claude Code sessions (!autocompact 150k|70|off).
• Fresh session per task — once !goal goes green and the thread is past 60k tokens, the agent writes its notes to project memory, then /clear (/new on codex/opencode) and re-reads them. !fresh now does it on request; !fresh off keeps threads.
• Prompt lint — "fix it", "still broken" are held for one tap: ✨ improve rewrites it with Claude Haiku in the style of your own prompts that landed first try, adding the error on the agent's screen and the !goal check as the done-condition; or send as is. !lint off turns it off.
• Corrections are counted — "no", "still same", "not working"… mark the previous task as missed. !coach shows what your first-try prompts have in common; !route stats shows the first-try rate per agent and task class, and auto routing (!route auto) uses it once an agent has 5 tasks in a class.
• Model ladder (!ladder on) — Claude takes light tasks on haiku, normal on sonnet, heavy on opus, and steps up one model when !goal fails the same way twice, the loop guard fires, or you correct it twice. It never drops to haiku on a context over 120k.
• Switches only when they are cheap — the prompt cache belongs to one model and one agent, so a switch re-reads the whole thread uncached. The ladder and !route auto change model or agent only when the thread is under 30k tokens (a new task, after !fresh); the one exception is a step up when the agent is struggling, at most once per task. !route stats shows what the switches cost.

Keep going until it passes

!goal npm test (or pytest -q, npm run build, anything with an exit code) makes the topic's agent prove it is done: after every finished turn nightmux runs the check in the project folder. Red, and the last 40 lines go back to the agent as its next prompt — fix the cause, don't touch the test. Green, and you get one ✅. The same failure three times in a row, or 6 rounds (!goal 10 npm test for more), and it stops and shows you instead of burning the night. Your next message resumes it; !goal off ends it.

Two servers

One bot, one group, topics spread over several machines. On the dashboard, servers → + add server shows four steps:

1. On the new machine: sudo apt install -y tmux git python3, then install Tailscale and sudo tailscale up into the same tailnet.
2. Install the agents you want there and sign each in once.
3. Paste the command the dashboard gives you — nightmux.py --join <code>. The code is good once, for 30 minutes. It fetches the bot settings and a fresh peer secret from this machine over the tailnet, writes the new machine's config as a peer, wires the Claude Code hooks and installs the service.
4. The new server appears on the dashboard. Pick it as the server in + new project, or send !server <name> in a topic to move it there (!server local brings it back).

How it works: Telegram lets one process poll a bot, so this machine stays the primary — it forwards each remote topic's messages to the peer that runs it, and the peer replies to Telegram itself with the same token. Peers talk on a listener that serves only /peer/ routes, with the shared secret. The dashboard and office show every machine's topics, metrics and limits; remove on a server card forgets it (once its topics are moved or closed).

The setup panel at the top of the dashboard lists anything not yet set up on this machine — bot rights, agents, tailnet access, gh, Chromium — each with the command that fixes it (the same checks as nightmux --doctor).

The dashboard

💬 on every topic card opens that topic as a chat: your messages, the agents' answers, their menus as buttons that work, a box to type in, and a "working…" line while the agent is busy. It reads new messages only, so the page never jumps. Add the dashboard to your phone's home screen and it opens like an app; 🔔 turns on notifications while it is open (closed-app push stays Telegram's job). The chat keeps the last 300 messages per topic in memory — Telegram keeps the full record.

+ new project opens a form: a name, an agent, a folder (defaults to projects_root/<name>), a server when you have peers, and an optional idea. It creates the Telegram topic, the folder and the session — with an idea, the agent gets the spec → build → ./check.sh brief and the !goal loop. On every topic card: the live agent and the rest of its bench as chips (tap one to make it live), + agent to start another beside it, and close, which stops the topic's agents, unbinds it and closes the Telegram topic — files are never deleted. The bot needs the Manage topics admin right to create and close topics. These actions require an X-Nightmux header, so another website you visit cannot trigger them.

http://127.0.0.1:<webhook_port>/ (or over tailscale serve) shows server metrics, limits per agent, every topic with a send box, and a chat analysis: per agent, how many prompts, model calls and tokens you spent, the cache hit rate, the average context each call carried, how often you nudged with "continue"/"yes" — and what to change to get better output for fewer tokens. The same report is !stats [days] in Telegram.

Share it

• !saved — what nightmux did for you, counted: tokens of context not re-read after compactions and fresh starts, turns worked between midnight and 7, queues resumed after limits, hand-offs, green checks, loops caught.
• !wrapped [days] — the same as a 1200×675 card with your agent stats, and a 🐦 button that opens a ready-to-post tweet.
• The night reel is signed with the repo link and gets the same 🐦 button.
• !public on — a read-only office at /public/<token>: agent states only, no screen text, prompts or buttons. Tailnet-only until !public expose, which puts just that path on the internet with Tailscale Funnel. !public off kills the link.

Night QA

!qa 03:00 http://localhost:3000 — every night the agent opens the app with the browser tool (added for you), uses it like a new user, and files each real bug as a GitHub issue labelled qa, with steps to reproduce. It does not touch the code; !issues picks the fixes up in the morning. !qa now, !qa off.

Leaderboard

!leaderboard shows the counts it would share — turns, night turns, limits survived, green checks, tokens saved, which agents; no project names, prompts or code. !leaderboard post adds them as a comment on the leaderboard issue with your own gh login, and the board ranks the crews. Nothing is sent without the post.

Race agents on one task

!race claude,codex add rate limiting to the login endpoint gives the same task to each agent, each in its own git worktree from your current commit (so nobody edits anybody else's files). When they are done you get one card: what each changed, whether your !goal check passes in its tree, how long it took — and a 🏆 button per agent. The winner's changes land staged in your folder; the other worktrees are deleted. !race diff codex shows one's diff first, !race cancel drops it all. Without a list it races the topic's bench (or the first two agents installed).

Voice notes

Send a voice note and, with a transcriber configured, the agent gets the words (shown back to you as 🎙 so you can check them). You choose where your voice goes — any command that takes the audio path last and prints text:

"transcribe_cmd": "sh -c 'ffmpeg -loglevel quiet -y -i \"$0\" -ar 16000 /tmp/nm.wav && whisper-cli -m ~/models/ggml-base.en.bin -nt -np -f /tmp/nm.wav'"

(whisper.cpp, fully local.) Without one, the agent gets the audio file's path.

Without Telegram

No bot? Press Enter at the token prompt in nightmux --setup (or leave token out of ~/.nightmux.json). nightmux runs dashboard-only: the dashboard at http://127.0.0.1:9090/ opens projects with + project, and each project's 💬 chat view (an installable PWA) is the channel — the same commands, buttons and replies you would get in a Telegram topic. Add a bot later by running --setup again.

The office

/office on the same port is your night crew as a pixel-art office: a room per topic, a desk per agent on its bench, every animation a real state — typing while it works (and what it's doing), hand up when it's asking (with the menu's real options as buttons), asleep with a countdown on a usage limit (with one-tap hand-off to another agent), an empty chair when its session is gone, sticky notes for queued prompts. Tap an agent to answer it, prompt it, or fail it over. Screen text is redacted before it leaves the machine.

Each agent has its own look — claude's hood, codex's cap, agy's headset, opencode's beanie — and the monitor shows what kind of work it is: code scrolling, a yes/no dialog, a limit bar, static. A finished turn sparkles; a hand-off flies a folder from one desk to the next. Every room has a lounge: idle agents get up and wander to the coffee machine, the TV or the toilet and come back; one that hit its usage limit goes to bed until the window resets. /office?demo plays the night shift above on a loop, no server state needed.

To open it from your phone without opening a port to the internet, put it on your tailnet:

tailscale serve --bg 9090          # https://<machine>.<tailnet>.ts.net

then "office_url": "https://<machine>.<tailnet>.ts.net/office" in the config, and !office posts it as a button in Telegram.

Plugins

Drop an executable file in ~/.nightmux-plugins/; its filename becomes a command. !weather runs ~/.nightmux-plugins/weather, its argument as $1, and whatever it prints to stdout is the reply — the same shape as the built-in !git/!grep. It never reaches a session's keyboard, so it works in a read-only topic same as any other read command. !plugins lists what's there.

How it works

Four files, no framework:

nightmux.py	the daemon: long-polls Telegram, watches tmux, everything above
nightmux_state.py	status-line sidecar — parks context %, limit windows and the transcript path where the daemon can read them
nightmux_stop.py	Stop hook — pushes the final answer as exact text, not scraped pixels
nightmux_notify.py	Notification hook — pushes permission prompts the instant they appear

The daemon reads the session's JSONL transcript when the sidecar is installed, which is why output arrives as clean text with a real tool trace. Without it, nightmux falls back to scraping tmux capture-pane — everything still works, just noisier and without the usage numbers.

A screen the classifier does not recognise is never guessed as idle — it is held as unknown and treated like a busy pane: a prompt sent to it queues instead of typing into a state nobody has confirmed is safe, with !raw to type it anyway.

One watcher thread polls every bound session; each topic gets its own worker thread, so a slow command in one topic never blocks another. Within a tick, each bound session's own tmux round-trips run in a small pool (up to 8 at once) so one slow or hung session cannot stall every other session's update behind it. The polling offset is only persisted past updates that have actually finished, so a crash replays work rather than dropping it.

ARCHITECTURE.md has the rest: threads, what survives a restart, how output is chosen, and the decisions that were rejected.

Run the tests: python3 nightmux.py --selfcheck (and the same flag on the three hook scripts). No framework, no fixtures — asserts that fail loudly.

nightmux --doctor triages an install without fixing anything: tmux found, config complete, token accepted by Telegram, hooks wired, service active — one ✓/✗ line each, exit 1 if anything is off.

python3 tests/test_panes.py runs the pane corpus: captured terminal screens and the state nightmux must read from each. Adding an agent whose TUI it misreads is one file — drop the pane in tests/panes/ as <what>.<busy|idle|waiting>.txt and the classifier is held to it from then on.

Config

~/.nightmux.json, mode 0600, written by setup:

{
  "token": "<from @BotFather>",
  "chat_id": -1001234567890,
  "allow_users": [123456789],
  "topics": {"12": "api"},
  "agent": "claude",
  "agents": {"opencode": ["opencode", "--continue"]},
  "autostart": {"api": "~/code/api"},
  "auto_restore": false,
  "projects_root": "~/code",
  "tz_offset": "Africa/Cairo",
  "autocompact": 70,
  "auto_continue": "continue",
  "modes": {"115": "readonly"},
  "auto_update": "1d",
  "failover": "codex",
  "poll": 2
}

agent is what !new starts. agents adds or overrides entries in the built-in table as [command, resume-flags] — those flags are the part most likely to drift as these CLIs change, so they are config, not code. autostart recreates named sessions after a reboot. For everything else a reboot killed, nightmux checks each bound topic against tmux at startup: with auto_restore it just relaunches and says so; without it, the topic gets a "machine restarted?" message with a Restore button instead of nightmux acting on its own — !restore does the same relaunch on demand. projects_root makes a new topic named after a directory start that project on its first message. !reload picks up hand edits without a restart.

failover names the agent a topic hands its held work to the moment its agent hits a usage limit — the working tree and the cut-off instruction go to codex (or whoever) instead of waiting hours for the reset. Without it, the limit message offers the same thing as one-tap buttons; !failover <agent> does it by hand.

auto_update (true = daily, or an interval like "12h"; off by default) runs each installed agent's own updater on that schedule — claude update, codex update, agy update, opencode upgrade — and posts what changed to the command-center topic, or General. !update [agent] does it on demand. update_cmds overrides or adds an agent's updater; "" turns one off. It is off by default because it runs installers unattended. Running sessions keep their version until relaunched: !kill yes, then !restore.

Requirements

Python 3.8+ (CI runs 3.8 through 3.13), tmux, a terminal coding agent, and Linux with systemd (the service is optional — python3 nightmux.py in a terminal works fine). Claude Code gets the hooks and the usage numbers; everything else runs on the terminal scrape.

Security

The bot token is a shell on your machine, and allow_users is the only thing between a stranger and your sessions. Read SECURITY.md before you add a second person or a second machine. It is short.

License

MIT. Changes are in CHANGELOG.md.