Amux

amux's pitch is narrow and honest: it is infrastructure for people who already have an agent subscription and a terminal habit. The mental model is a fleet, not a chatbot. You register sessions (`amux register myproject --dir ~/Dev/myproject --yolo`), start them (`amux start myproject`), serve the dashboard (`amux serve` on port 8824), and steer from there. The pieces that justify the layer are the ones a single agent never needed. Claiming is compare-and-swap in SQLite, so two workers cannot take the same card. Closing a card as `done` requires evidence of what actually ran; `verified` requires a peer check. Dependencies are expressed as `blocked_on`, and `GET /api/board/ready` returns only tasks whose dependencies have cleared. Workers talk directly over a REST API and the server stamps the true origin on every message — the vendor's claim is that provenance is a fact, not a claim, which is the kind of thing that matters when you are reconstructing why an overnight run went sideways. Recovery is the second load-bearing piece. The watchdog monitors context, auto-compacts at a threshold, restarts crashed sessions, and replays the last message, so a 3am crash does not mean dead terminals and a lost task. The message ledger keeps every prompt, reply, and delivery as a fetchable row. Schedules accept plain English rather than cron syntax, and autonomous loops let an agent pace itself: schedule the next wakeup, drain the ready queue, go quiet when there is nothing to do. The vendor cites an overnight run of 7 cards closed with 0 interventions. The July/August 2026 changelog entries address the thing fleets actually break on — observability. The 2026-07-31 plan strip puts the agent's live task list and next steps inside the peek panel, so you can see intent without attaching. The 2026-08-02 review dashboard adds a Trends view and a fleet org chart that groups activity by theme and surfaces blockers first, which is the closest thing here to a morning standup. Where it doesn't fit: the self-hosted path wants Python 3.8+, tmux, and an installed and authenticated Claude Code CLI, and sessions literally run Claude Code inside tmux panes. If you run one agent, you gain nothing over a plain terminal. There is no SSO or RBAC in what we can see, which rules it out for teams with a compliance gate. And the workflow is tmux-shaped — teams standardized on editor-integrated agents will bounce off it. Cloud ($20/mo, or $200/yr billed annually) removes the server but not the BYO Anthropic API key, so your token spend stays separate.