49 Agents IDE

You will like this if you already run three or four agent CLIs at once and have lost track of which window is which. The deciding trait is the zoomable canvas: panes keep their place in space, so you navigate by memory of where a thing was rather than by cycling tabs. Zoom out, see every project on every machine at once.

It is wrong for you if one repository and one agent is your whole day, because the canvas is overhead you will not pay off. Pick it when you are juggling machines. Pick Agent Deck if a single terminal manager covers you.

The risk is blast radius. Every pane is a real shell with your own config, and broadcast mode types the same keystrokes into several terminals at the same time. One careless line goes to every session, on every machine attached to the canvas. There is no layer between an agent and the box it landed on, and nothing documented that scopes a broadcast to a subset.

What it gets right is the honesty of the surface. Scrollback is real, colour is real, and nothing is re-rendered through a wrapper that loses a character at the wrong moment.

1. The layer asserts no intelligence of its own. Sessions are served as ttyd-backed terminals, so each agent's loop is untouched and what is added is display. 2. Verification is delegated to two artefacts placed beside the session, a git graph and an editor pane, which is a defensible separation of production from inspection.

2. No evaluation is published, and none is owed, since no capability is asserted. The one claim worth recording is architectural: terminal traffic is not retained server-side, which constrains what a relay can leak.

551 stars and no invoice anywhere. There is no revenue line to model, which turns the survival question into a question about one maintainer's attention. The obvious business is the relay, since somebody has to host the thing that connects machines, and today the project tells you to host that yourself.

Moat: none a competitor could not rebuild in a quarter, because the value is arrangement rather than capability. Likely acquirer: nobody, though the canvas idea gets copied by whoever ships the next agent dashboard. Position: use it, expect to maintain it.

Zero per seat, which ends the finance conversation and starts the operations one. There is no identity story in the row: no SSO, no SCIM, no audit trail, and access is whatever Tailscale or the LAN grants. Sixty engineers sharing one surface means sixty shells reachable from a browser.

It does not run headless, so nothing here becomes a pipeline step I can measure or gate. Onboarding is an afternoon for anyone who already lives in tmux and longer for everyone else. Not yet.

BUSL-1.1 is not open source and I will say so, but the additional use grant covers me and the whole thing converts to MIT on 2030-02-26, which is a date I can plan around. Until then I read every line, and my key stays my key, because the agents in the panes are the ones I already pay for.

What I do not get is MCP. There is no client and no server, so the tool servers I run do not attach here. This is a window manager for other people's agents, not a place to hang capability. Local weights are not part of the deal either.