Roadmap
The aim: your agents run on your own machines, keep running whatever happens to the window, and you can see, answer and direct them from wherever you are, without Alto becoming one more dashboard to tend.
The terminal itself; sessions outliving the window; several clients per session; the background service and keep-awake; upgrades that keep sessions, programs still running (alto reload); agent status, marks and notifications; coming back after a reboot; reaching other devices from the command line, in tabs and in Settings, with no pairing for your own devices; slow links; working beside an agent (work panes, alto run, who’s typing, alto mcp).
Your own fleet, from the command line: joining by Tailscale tag; setting up a server with nobody there; what needs you on every machine; what each machine is (alto devices); putting work where it fits (alto new --in, pools, drain); sending your code with a session and fetching its branch back (--repo, alto fetch); every session on every machine in one list, grouped by repository, agent, OS or kind, each named as machine:id; alto run on Windows through PowerShell.
Who may do what: rules for other devices and for agents (access.toml, your organization’s managed settings, your tailnet’s policy), checked on every request and as sessions are typed into, with alto access check to ask; joining a session as an observer, a peer or a moderator, with who’s watching in every listing; agents that type only while someone watches; and a signed log on each device of who did what there (alto log, alto log verify, alto log export --ocsf for a security team’s tools). See Access and The log.
Moving around the fleet: the switcher (Cmd-P) over every session on every machine, needing you first; Cmd-Shift-J to the next session needing you; the fleet view (alto fleet, Cmd-Shift-L), grouping and acting on many at once, by keyboard and mouse, in a tab or any terminal. Adding a machine from here over ssh (alto add HOST).
macOS is used daily. Linux builds and passes every test; its window is checked less. Windows builds and passes its tests in CI on a real machine, with window checks by hand.
Installing in one line Planned · Phase 9
Section titled “Installing in one line Planned · Phase 9”curl … | sh and its PowerShell twin, from published builds; Homebrew and Scoop, planned too; WSL as a machine of its own; credentials handed to a session when it starts.
Machines that start when work needs them Planned · Phase 9
Section titled “Machines that start when work needs them Planned · Phase 9”--in os:linux with no Linux machine running starts a container that joins by itself, runs the session and goes away when it’s done. Docker first, then other providers as programs of your own.
Your phone as a real client In progress · Phase 6
Section titled “Your phone as a real client In progress · Phase 6”Native iPhone and Android apps over the same Rust core and protocol. See Remote access with Tailscale.
Agents that run agents Planned · Phase 7
Section titled “Agents that run agents Planned · Phase 7”A boss is just a session whose agent has tools to start other sessions, hand them tasks and hear back; a boss can have bosses under it, and its workers are placed across the fleet as --in places sessions. The work is a tree of tasks kept on disk as markdown, one writer per file (the boss writes the brief, the worker its progress and result), so nothing needs merging and everything survives a restart.
Alto never plans or splits up work. It keeps the record, wakes an idle agent when a message arrives, rolls progress up the tree (“3 of 5 done, 1 needs you”), and enforces who may read, type into, start or stop which session. Permission prompts stay with you unless you hand them to a boss.
- Shared machines Planned · Phase 9: several people and agent roles on one machine, each as its own account, under rules, with a log.
- Extensions Planned · Phase 11: small programs, sandboxed, that add what travels between machines with a session, and commands of their own.
- Agents in a box, per session Planned · Phase 7: a session fenced by the operating system’s own sandbox (files and network limited, starting instantly), or in a small virtual machine, chosen per session or per role.
- Updates you don’t notice Planned · Phase 8: the app updates itself and the service follows. Sessions already move to the new build in place.
- Packages Soon: Homebrew first.
- Open source Planned · at launch: AGPL-3.0, with commercial terms for those who need them.
Not planned
Section titled “Not planned”Kept deliberately small: no repo, worktree or browser panels; no planning or scheduling logic of its own (agents do that); no syncing with issue trackers. Any of these comes only if real use shows a need.
If needed later: a relay for reaching sessions without Tailscale, attaching to tmux on servers that only have tmux, and a tmux stand-in for agent tools that open panes with tmux commands.