Create
Register a digest-pinned OCI image under a stable instance name.
barista create agent-42 …
Release the sandbox. Keep the session. Memory, disk, and working context stay together while compute rests. Resume returns the same process—not a reconstruction.
-
Workingturn 3 · context live
-
Restingmemory + disk retained
-
Released0 CPU · 0 host RAM
-
Wake firedscheduled alarm
-
Runningsame memory · same digest
Illustrative lifecycle.
The lifecycle
One named session. Four deliberate moves.
Address work by name and let Barista manage its runtime lifecycle. The service language is light; the guarantees underneath are exact.
Run
Exec, transfer files, inspect readiness, and keep useful work in place.
barista exec agent-42
Pause
Capture memory and disk on a capable runtime, then release the sandbox.
barista pause agent-42
Resume
Restore the process, reseed entropy, step time, run hooks, and continue.
barista resume agent-42
Where it fits
Made for work that accumulates context.
Named, long-lived sessions whose useful state should survive while compute rests.
Cloud agent harnesses
Keep Codex, Claude, ACP, and other interactive agent harnesses ready between turns without rebuilding their working context.
Long-running agents and workers
Let agents and workers rest between tasks, approvals, or scheduled runs, then continue with the same memory, disk, and process state.
Stateful online services
Wake a named dev server, preview environment, or session-affine app when traffic arrives once the request gateway ships. Not for static sites or general stateless hosting.
Available now
The core session loop ships.
Developer previewThe reference agent kept the same conversation across five consecutive 60-second pauses. Median Resume to first JSON-RPC response was 368 ms on the measured setup; while fully paused, the session held zero CPU and zero host RAM.
Memory-preserving pause and resume
Exact-memory restoration on supported hypeman backends; unsupported hosts report the capability as absent.
Lifecycle and CLI
Bring any digest-pinned OCI image with no workload SDK, then create, start, exec, pause, resume, inspect, and destroy through scriptable barista --json output.
Named, reusable snapshots
Retain a point-in-time snapshot, restore its exact bytes more than once, and remove it explicitly.
Scheduled wake
Give one paused session an alarm and let the journaled lifecycle wake it at the requested time.
Bucket coordination protocol and fleet CLI
Conditional ownership and name resolution are implemented. Node-side fleet membership is not wired yet, so multi-node operation is not currently a mode you can turn on. Read the fleet status.
Ways to wake
Today
01
Today
02
Soon
03
The moment can come three ways.
Barista owns the return path. Two wake edges work today; the gateway that turns an incoming request into a transparent wake is coming soon.
On command
The explicit operator path and the machinery beneath every other wake. Use it in scripts, tests, and recovery procedures.
barista resume agent-42On schedule
Persist an absolute alarm with the session. When it falls due, Barista submits one idempotent resume even across a node-agent restart.
barista wake-at agent-42 2026-08-09T09:00:00ZOn incoming request
The planned gateway will resolve the session name, hold the request while the owner restores, and forward it once the workload is ready. It is roadmap—not a current claim.
What “ready” means
The boundaries stay visible.
One API spans hosts with very different guarantees. Barista reports the truth before accepting work and keeps a weaker result from passing as silent success.
Memory preservation is a capability, not a label
A runtime may claim memory snapshots only when an in-memory counter continues and the guest shows no reboot. Otherwise the capability is absent and operations that require it fail explicitly.
Warm state is node-local today
Losing a node loses its local memory snapshots. Another owner can cold-boot from desired state with a degradation event, but the remote snapshot tier needed for warm cross-host recovery remains later work.
Live checkpoint is not approximated
The rank-1 substrate freezes a running workload while creating a snapshot. Checkpoint therefore fails with CAPABILITY_MISSING instead of pausing and calling that result live.
The fleet protocol is ahead of fleet membership
The bucket protocol and fleet CLI are real and tested. A node cannot safely join that fleet yet; durable ownership recovery and self-fencing are being designed before the wiring is enabled.
macOS pause works; host reachability does not
Memory pause and resume work on Apple Silicon, but upstream hypeman issue #358 prevents the macOS host from reaching the guest network. The complete host-to-session scenario remains Linux-only.
Egress policy is capability-gated
A mediated policy is accepted only where the runtime reports egress_control. Unsupported runtimes refuse creation with CAPABILITY_MISSING rather than starting with unrestricted traffic.
Open source · developer preview
Keep the work ready. Let the machine rest.
Follow Barista for implementation progress, measured results, and the next wake edge.