Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.cyberun.cloud/llms.txt

Use this file to discover all available pages before exploring further.

A Cyberun agent is a process you run on a machine with a GPU. The agent maintains an outbound connection to the Cyberun gateway, identifies itself with an agent credential (ak-...), and waits to be dispatched a task. Your machine never accepts inbound connections. You bring the hardware. Cloud schedules the work.

​
How an agent connects

The agent dials out to Cyberun and holds the connection open. Your machine never accepts inbound connections — you don’t need to open a port or change a firewall. As long as the GPU machine has ordinary outbound internet access, it can join the team.

​
Status

In the Agents sidebar page, each agent shows one of:
StatusMeaning
idleConnected, no task running. Ready to dispatch.
busyRunning a task. Other tasks queued behind it.
syncingPulling models or node packages on first run; not yet eligible.
tunnelA ComfyUI web tunnel is open for interactive use. Won’t accept tasks until the tunnel closes.
unhealthyConnected, but its ComfyUI health check failed — it won’t receive tasks until it recovers.
If the agent disappears from the list entirely, its connection to the gateway dropped. Check that the agent’s system service is still running on the GPU machine (the installer registers it under a name like cyberun-agent-comfyui). See Connect an agent for the per-OS verification commands.

​
Labels

Labels are how workflows pick eligible agents. Common defaults:
  • gpu — the agent has a usable GPU.
  • comfyui — the agent runs the ComfyUI runtime.
  • nerfstudio — the agent runs the Nerfstudio runtime.
Most labels are set automatically. The installer derives them from the --tool you chose and from the detected hardware; the agent advertises them when it connects. You can add extra labels in the agent’s local config file (under agent_labels: in cyberun-agent.yaml) — useful for tagging a specific machine for priority dispatch, e.g. team-a, gpu-h100. A workflow declares the labels it requires. Tasks dispatch only to agents whose label set is a superset of the workflow’s requirements. A workflow asking for gpu, comfyui runs on any agent advertising at least those two. Labels are case-sensitive. Don’t ship the same logical capability under two names (comfyui vs comfy) — pick one across your team.

​
Health and resource reporting

The dashboard shows per-agent telemetry on the Agents page:
  • GPU model and VRAM usage
  • Runtime version (for example ComfyUI commit, Nerfstudio version)
  • Last heartbeat timestamp
  • Current task ID
When the resource is fetched via an integration key (sk-) instead of a signed-in user session, some host-level stats appear in a reduced form — this is by design, so a leaked key reveals less about your team’s hardware than a user session would. Sign in through the dashboard if you need the full telemetry.

​
Remote agent tools

Some deployments enable a remote-access toolkit on the agent detail page, with three sub-tabs:
  • Terminal — interactive shell into the agent host.
  • Files — browse, upload, download, rename, copy, delete, mkdir.
  • Desktop — view the agent’s screen and send mouse, wheel, and keyboard input.
These tabs use short-lived tunnel credentials and are scoped to the team. Treat them like SSH — only admins should use them, and only on agents you trust. If the tabs aren’t there, this deployment doesn’t include remote access. Contact your team admin if you need it.