workspace (a workspace ID) and environment (the name of the Runcrate environment to act in; it defaults to the one the connection was signed in to). Values for a launch are never called environment: run_job takes them as variables.
Instances
Guardrail parameters on create_instance
Both are optional and off when omitted. Ask your agent: “Launch an H100, stop it after 4 hours or after 30 idle minutes.” See Spend guardrails.
Clusters
A cluster is several GPU machines on one private network, with optional Slurm, Kubernetes or Ray, bought by the machine. It is all or nothing: every machine starts or none does, billing starts when every machine is up, and terminating it stops every machine together.Jobs
A job runs one command on its own GPU machine and stops the machine, and billing, when the command exits.Schedules
A schedule starts an instance or a job later: once atrun_at, or repeating on a 5-field cron in a timezone. Runs start within about 2 minutes of their time, as whoever set the schedule up, through the agent sign-in it was set up with: ending that sign-in pauses the schedule. Keep tokens and passwords in the environment’s secrets (set in the dashboard), never in launch_script or startup_commands, which everyone in the environment can read. Every schedule tool takes workspace and environment (the environment name) to choose where it acts.
Secrets
Secrets are environment variables every instance and job in an environment gets at boot. An agent can see their names, so it knows a command can read$HF_TOKEN, but never their values.
No MCP tool adds, changes or removes a secret. A workspace owner or admin sets them in the dashboard, so a value never passes through the agent.
Firewall
The firewall decides which inbound ports the machines of an environment accept, and from where. The tools act on the environment the agent is connected to; only workspace owners and admins can change rules.HTTPS links
An HTTPS link serves one port of an instance athttps://<port>-<instance>.runcrate.app. Workspace owners and admins, and whoever launched the instance, can add and remove links.
Feedback on model answers
Score an answer your agent got from the inference API, by its request id. The score shows with the request in Models → Observability.Storage
Billing
No MCP tool sets or changes a budget. Budgets are set in the dashboard by a workspace owner or admin, so an agent cannot raise its own ceiling. See Monthly budgets. Likewise,
top_up only works once an owner or manager turns on the billing permission for the sign-in, and only up to its daily and monthly caps.Model calls
See Observability for sending end users, sessions and tags with each call.