Limits

Latency

Desktop methods run synchronously over authenticated HTTP RPC. Typical round-trip from the MIOSA edge:

For high-frequency loops, prefer bash() for multi-step operations over many individual GUI calls. GET /desktop/window-state returns the screenshot and the accessibility tree together, so perception costs one round trip instead of two.

Coordinate space

Coordinates are screen pixels with the origin at the top-left, matching what GET /desktop/screen-size and GET /desktop/cursor report. A vision model may be shown a scaled-down screenshot; map its output back to screen pixels before sending an action.

Screenshots

Screenshots are PNG. A full desktop image is typically a few hundred kilobytes; regions are smaller. Use /desktop/screenshot/region when only part of the screen matters, and window-state when you need pixels and structure at once.

Sizing

GPU support for accelerated desktop and GPU-aware workflows is plan-dependent.

Lifecycle

Stopping a Computer pauses compute and preserves the filesystem and browser profile. Restarting resumes from that state. Desktop endpoints only accept requests while the Computer is running; otherwise they return 409 COMPUTER_NOT_RUNNING.

Authentication

Desktop methods require the same workspace API key as every other MIOSA resource. Call them server-side. For browser-side control, mint a scoped token instead of shipping the API key - see Embedding & streaming.

Rate limits

Rate limits vary by key, tenant policy, and endpoint family. Honor X-RateLimit-* and Retry-After response headers. Individual desktop actions are fast, but a tight perceive-act loop still issues many requests; batch where the work does not need pixels.

Errors