fleet-status

Fleet status

Every machine runs the hwprobe-fleet agent, which reports CPU %, RAM, disk, GPU and the Claude Code accounts signed in there (5-hour and 7-day usage windows). A central hub aggregates the reports and exposes them over MCP and HTTP.

Read it

Prefer, in order:

1. MCP tools if your session has them: pick_host (ranked recommendation), fleet_status (raw numbers), claude_usage (accounts + windows), host_stats(hostname), fleet_table. They come from the fleet-hub MCP server (HTTP) or the local hwprobe MCP server.
2. CLI when installed on this machine:

   hwprobe-fleet                 # table of every host
   hwprobe-fleet pick --gpu      # ranked hosts; --ram 16 --disk 50 --max-claude 80 also work
   hwprobe-fleet status --json   # full document
   

3. HTTP from anything else (curl, a browser, an Orca panel): GET http://<hub>:7790/v1/fleet with Authorization: Bearer <token>, or on any machine with the agent GET http://<host>:7788/v1/fleet (polls peers; no token on the tailnet). The dashboard is the same URL without /v1/fleet.

Decide

• status must be ok; stale, lost, no-agent, offline mean the numbers are not current — do not delegate there without checking.
• Rule out a host when: cpu.percent > 85, memory.available_gb is below what the task needs, disks.primary.free_gb is below what the task writes, or the task needs a GPU and gpu.gpus is empty / utilization_percent is high.
• Claude usage: claude_accounts[].session_percent is the 5-hour window, weekly_percent the 7-day window, scoped per-model windows (e.g. Fable). Near 90 % on every account on a host means a Claude session there will throttle soon; pick a host (or account) with headroom. Reset times are in session_resets_at / weekly_resets_at.
• Map a fleet hostname to the Orca target with the same name (orca host list, orca environment list) or the SSH alias of that machine.
• A machine missing from the list has no agent: say so instead of assuming it is free.