Know the moment it stops.
Your agent works while you look away. almost tells you the second it finishes, stalls, or needs a decision: on your desk, and in your team's channels.
Claude Code, Codex, Gemini CLI, or anything that can run a command.
- refactor auth middlewareclauderunning2:14
- migrate notices schemacodexneeds you0:42
- backfill 23 state parsersclauderunning8:32
- write e2e coveragegeminifinished6:03
Your agent is not the slow part.
Every stretch between the last token and your next glance is time you paid for and did not get. Multiply it by every agent you run, and every person on the team running one, and it stops being a rounding error.
- It finishes
- The last token lands and the turn closes. You get the banner before you have finished reading the other tab.
- Stop
- It needs you
- A permission prompt, a destructive command, an ambiguous instruction. The run is blocked and nobody is watching it.
- Notification
- It goes quiet
- No output, no error, no exit. The failure mode nothing else reports, because nothing happened.
- Idle
One signal, everywhere you already are.
The desktop banner is on by default and needs no account. Add a channel and the same events reach the rest of the team, so a blocked run is somebody's problem in seconds rather than at standup.
- SlackIncoming webhook
- DiscordIncoming webhook
- TelegramBot token
- WebhookAny JSON POST
An agent finishing is not the work being finished.
Each project gets a board, and the agents move their own cards across it. A run that starts goes to Running, one that hits a permission prompt goes to Blocked, and one that stops goes to Almost rather than straight to Done, because stopping is not the same as passing. Whatever your real gate is, CI, a merge, a review bot, send it a done and the card closes itself.
- BacklogNot startedadded by you
- RunningAn agent is on iton start
- BlockedWaiting on a personon notification
- AlmostAgent stopped, not verified yeton stop
- DoneChecks passed or mergedon done
A signal, not a logger.
Your agent hands the hook a path to the entire conversation. almost drops it and has no code that reads a transcript. By default the agent's own words never leave the machine: a permission prompt becomes a category, not a quote.
What the agent said
“needs your permission to run: rm -rf /etc/secrets”
What almost sends
“permission required”
You do not have to take our word for it. Run almost preview and it prints the exact request body, then sends nothing.