Cloud|

Operate Studio machine execution

Monitor bounded Studio executions and their machine audit records.

2 min read Updated 2026-10-23 #operations#studio#service-accounts

Studio supports explicit machine execution through the existing Assistant service and its managed Chromium sandbox. Deploy the Assistant migration and runtime assets together. No new service, environment variable, personal account or default execution grant is required.

Existing app grants retain their meaning. Machine execution and effect rights start empty. App managers configure them in Studio permissions. Machine secrets are separate from personal secrets and are encrypted with the existing Cloud secret service.

Executions share Studio's existing limit of 8 code hosts and its 45-second operation deadline. Database, shared storage and HTTP use their existing configuration and budgets. Machine execution does not create or configure an app database; an app manager prepares it first. Storage, database and HTTP operations observe both execution cancellation and cancellation of their individual runtime request.

Shared file storage counts payload bytes and each file's stored key/media-type metadata record against assistant.storage_total_mib. This applies to all Studio storage callers, including existing files after upgrade. No migration or new setting is required. Payload sizes shown in listings stay unchanged. Existing files remain readable; replacements that do not increase usage and deletions remain available when usage exceeds a lowered total budget. See Studio file budgets.

Audit events named studio.machine.* identify the machine account, credential reference, app, published version, operation, request correlation and outcome. They contain no source, input bodies, secret values or output payloads. OAuth references are one-way credential fingerprints; API-token references are credential IDs.

Execution status records retain the existing 2-day HTTP claim window and are cleaned at admission. Requests outside the 1-day admission window cannot replay after cleanup. An abandoned running request is reported as EXECUTION_UNKNOWN when inspected after its operation deadline.

When execution ends without a known result, inspect the app's shared data and external systems before issuing a new request. Studio never automatically replays an admitted request. Revocation prevents the next operation; it cannot undo an external effect already admitted.

Connect a coding agent

The Cloud skill provides compact working instructions. MCP supplies exact current documentation when details matter.

CLI command
bunx skills add https://docs.example.com

This command installs the working instructions published by this website in the selected coding agent.

For exact, current details, also connect the MCP server through one of the agent tabs.

Installation uses the open-source Vercel Skills CLI.

The skill and MCP complement each other: the skill describes workflows, while MCP supplies current documentation.