Features

Everything a sandbox should do. Nothing it shouldn't.

Shells wraps Omega's production sandbox workload in a deliberately small surface: create with a runtime, control the lifecycle, watch every action land in the terminal, and tear it all down without residue.

Five language presets

bash, node, python, rust, go — toolchain pre-installed, ready at create.

Three isolation classes

wasm for speed, container for familiarity, microvm for hard boundaries.

Multi-region placement

Shells run close to your callers; each row shows its region in the fleet.

Per-second metering

Sessions are metered when they start, not when you remember to stop them.

Live fleet

The table polls Omega continuously — status changes appear without a refresh.

⌘K everything

Run, refresh, or open any shell from the command palette.

Create

Runtime is a decision, not a default.

The create dialog asks three questions that matter: which language toolchain, how hard the isolation boundary, and whether the network exists at all. Sealed is the default — egress is something you grant, not something you forget.

Create-shell dialog: language, runtime class, and network mode
Shell detail sheet with the always-dark activity terminal and danger zone

Lifecycle

The terminal remembers.

Restart, snapshot, fetch logs, reinstall the OS — every action you or your agents take is appended to the activity terminal with a timestamp. Stop and wipe_state live apart in a danger zone with typed confirmation. Logs are pulled on demand through the same action API, so what you see is what actually ran.

Credentials

Secrets with a half-life.

Term credentials connect you to a live session — and that's all they do. Minted on demand, masked in the console, expired in minutes. The activity terminal records that a mint happened, never the material itself.

Term credentials dialog with masked token and expiry notice
Command palette with run, refresh, and per-shell navigation

Speed of use

Keyboard-distance to everything.

⌘K opens the palette: run a new shell, refresh the fleet from Omega, or jump into any instance by name. The whole console is keyboard-navigable — built for people who live in terminals.

For agents

The same product, machine-shaped.

Four MCP tools mirror the console exactly — list, create, action, term. Agents operate within the same Keystone boundary as humans, with one extra gate: the shells.agent.invoke entitlement.

shells-mcp-server — tools/list
shells_list — list shells visible to the caller
shells_create — create with language / runtime / network
shells_action — stop · restart · reinstall_os · snapshot · wipe_state · logs
shells_term — mint short-lived live-session credentials