The centralization promise is 'every agent, every machine, one inbox.' The plumbing exists — apps.hostlink is multi-host, each machine runs a host-agent daemon that enrolls with signed WS — but the aggregated experience isn't proven across multiple machines running different providers at once.
Scope
- Multi-host: Mac + Windows + Linux VPS enrolled to one backend over the private network (Tailscale/headscale). Sessions from every host appear in one forum, distinguishable per-host (topic prefix or label). Verify a daemon restart on one host doesn't disturb others.
- Multi-provider drive engines: today the interactive engine drives
claude -p. Generalize to a AgentRuntime interface so Codex (codex exec), Gemini CLI, and others can be driven — not just observed via MCP. One session, correct engine, selected by the thread's provider.
- Windows host-agent: the interactive engine is pure stdio (no tmux), so Windows is unblocked — package the daemon + a Task Scheduler/NSSM supervisor mirroring
deploy/orc-stack/.
Done when
Two machines with two different providers (e.g. Mac/Claude + Windows/Codex) drive concurrently into one chat inbox, each session labelled by host+provider, surviving a per-host daemon restart.
The centralization promise is 'every agent, every machine, one inbox.' The plumbing exists —
apps.hostlinkis multi-host, each machine runs ahost-agentdaemon that enrolls with signed WS — but the aggregated experience isn't proven across multiple machines running different providers at once.Scope
claude -p. Generalize to aAgentRuntimeinterface so Codex (codex exec), Gemini CLI, and others can be driven — not just observed via MCP. One session, correct engine, selected by the thread's provider.deploy/orc-stack/.Done when
Two machines with two different providers (e.g. Mac/Claude + Windows/Codex) drive concurrently into one chat inbox, each session labelled by host+provider, surviving a per-host daemon restart.