Fleet Roster¶
Add/update or clear a machine's entry in the inference.fleet roster (config.local.yaml) — its role label and optional hardware specs. Enriches live-discovered fleet machines, joined by device_id.
MCP name: fleet_roster
Category: inference
Parameters¶
| Name | Type | Required | Description |
|---|---|---|---|
action |
str |
No | "set" (add/update) or "remove" (clear the entry). Default "set". |
device_id |
str |
Yes | The machine's provider device identifier (the join key). Required. |
gpus |
list |
No | List of {name, vram_gb} for a peer — machines can have several GPUs. Omit to leave unchanged; pass [] to clear. (The local machine reports its own live.) |
ram_gb |
float |
No | Optional system RAM in GB for a peer. |
role |
str |
No | Human label for the machine (e.g. "Remote compute node"). The primary field. |
Details¶
Writes a machine's entry in the inference.fleet roster (persisted to
config.local.yaml). The roster only enriches live-discovered fleet machines —
joined by device_id — so device_id is the only required field; role is the
primary human label and hardware specs are optional (peer hardware can't be
auto-detected; the local machine reports its own live, so its roster specs are
ignored for display).
action: "set" adds or updates the entry; action: "remove" clears it (the
machine still appears via live discovery). Backs the inline editor in the
dashboard's Settings › Inference fleet section; the dashboard route
(POST /api/fleet/roster) wraps this capability, gating on read-only mode and
busting the fleet cache + publishing fleet.changed on success.
Returns {success, action, device_id, note}, or {success: false, errors_by_field}
on validation failure (e.g. a non-numeric VRAM value) so the form can highlight
the offending input.