tlive usage guide
tlive is a self-hosted monitoring/approval layer. Claude Code sessions report
through global hooks; Codex sessions are watched through an app-server
companion process (no hooks, no trust step). Completions and failures land in
IM (Telegram/Feishu) and the web dashboard, where you can reply-to-continue.
The posture (tlive mode, default notify) decides whether approvals are
held for a remote answer, in escalation order: off makes every hook a no-op;
notify only watches + notifies (the shim never holds an approval — prompts
stay 100% native) and reports to the machine (desktop toast + dashboard)
— IM stays quiet about a dialog only the terminal can answer, except once
ever per chat, when a card explains why and offers the switch to full;
full is the posture that puts approvals on your phone: it turns on
remote approval for the main session (Allow/Deny from IM/desktop/dashboard),
in parallel with the terminal dialog — first answer wins; all holds
sub-agent approvals too, with no terminal dialog for a held one until the
window ends, so use it only when nobody is at the keyboard (tlive mode full
goes back). Also settable from IM: /mode (a bare /mode replies with the
ladder). The daemon auto-starts with new sessions (disable via
daemon.autoStart: false).
Commands
tlive setup— configure IM credentials + register the Claude/Codex plugins (hooks ride the Claude plugin; Codex needs none).--hooks-onlyre-registers plugins only; add--claude/--codexto pick a vendor.tlive status— daemon health, effectivemode, channels, and the Codex companion state (running/degraded/off; degraded or off = Codex approvals local-only).tlive mode off|notify|full|all— set posture (see intro). Persisted to config, takes effect on the next hook;notifyis the default,full= remote approval on,all= also holds sub-agent approvals.tlive run <cmd>— wrap a process: local terminal + live web terminal (QR to open).tlive url— print the dashboard link + QR code.tlive logs -f— follow the daemon log.tlive start/tlive stop— explicit lifecycle (start is rarely needed; sessions lazy-start the daemon unless autoStart is off).
Diagnostics
- No IM messages:
tlive statusfor channel config;tlive logs -ffor send errors; confirm the daemon is up after starting a session. - Codex has no remote cards: check
tlive status— the companion line must sayrunning.offmeans codex isn't on PATH (or Windows);degradedmeans codex is there but nothing is answering on the app-server socket — see~/.tlive/codex-appserver.log. Neither is permanent: tlive keeps checking, so installing or reinstalling codex recovers on its own with no tlive restart. Either way Codex still prompts locally; nothing is ever auto-run. - No approval card ever arrives: check
tlive status— themode:line must sayfullorall. The defaultnotifynever sends approval cards (tool prompts stay local); enable remote approval withtlive mode full. - Claude approval card unanswered (in
fullorall): the local dialog stays live the whole time (parallel channels, first answer wins); answering locally resolves the remote card as "answered in terminal". The remote window defaults to ~24h (approvals.windowSec, shared by both vendors). - Web page unreachable:
tlive urlfor the current link (token is in the URL); phones need the same LAN (or your own reverse proxy/VPN — tlive has nopublicUrlconfig, and cards never carry the link).
Security model in one breath
- Never auto-allow: unanswered → Claude's local dialog governs / Codex's native prompt governs. Deny always carries a reason.
- Read-only tools (Read/Glob/Grep) pass by default.
/safe onalso auto-allows routine ops (non-dangerous Bash, non-sensitive edits) — the danger floor (rm -rf, sudo, .env/.ssh writes…) still asks and no config can lower it./trust onpauses approvals entirely (high risk — pair with allowedSenders). - Runtime switches flip the same state from either entrance: IM commands
(/mute /trust /safe on|off, /mode for the posture ladder) and the CLI
(
tlive mute|trust|safe on|off,tlive mode off|notify|full|all)./mute on= go quiet; it silences IM notifications ONLY. The desktop toast is independent of/muteand has no on/off switch of its own: one fires per thing that starts waiting on you (a pending approval, or the idle "waiting for your input" nudge), named with its project and call, and tlive never touches it again — it ages into the notification centre like any other app's, a log rather than live status (the dashboard is the live view). Silence it with your OS's Do Not Disturb, ortlive mode offto stop tlive entirely. A finished turn notifies the desktop too, from the Stop hook and after the same grace as the IM card; a failed tool does not, it blocks nobody. - Vendor-side
permissions.denyalways wins; tlive never overrides it.
First-time onboarding
When the user says "help me set up tlive" (or runs /tlive:setup), walk them through:
tlive statusto check the engine; missing →npm i -g tlive.- No channels → collect Telegram (bot token + chat id) or Feishu
(appId + appSecret) credentials and merge into
~/.tlive/config.json:{ "allowedSenders": [], "adapters": { "telegram": { "token": "…", "chatIdAllowList": ["…"] }, "feishu": { "appId": "…", "appSecret": "…" } } } tlive start→tlive statusto verify channels;tlive urlfor the dashboard.- Offer remote approval: tlive defaults to
notify(watch + notify only). If the user wants to Allow/Deny tool calls from their phone, runtlive mode full(holds each tool call for a remote answer; reversible withtlive mode notify). Leave it innotifyif they only want monitoring. If they say they're stepping away and want sub-agent approvals on their phone too, that'stlive mode all— flag the trade plainly: a held sub-agent has no terminal dialog until the window ends, so it only pays off when nobody is at the keyboard (tlive mode fullto come back). - Codex needs no extra step — the companion connects as soon as an app-server
is answering, and tlive starts one if none is. If status says
off/degraded, that's diagnostic info, not a setup task, and it clears itself once codex is installed — don't tell them to restart tlive.
