kestrel-cloud-gcp 0.2.0

GCP Compute Engine provider for Kestrel Sovereign — provision GPU instances, run training jobs, lifecycle management

kestrel-cloud-gcp

GCP Compute Engine provider for Kestrel Sovereign agents. Provision GPU instances, run LoRA training jobs over SSH, and manage instance lifecycle from your agent's tools.

Installation

uv pip install kestrel-cloud-gcp

The feature is auto-discovered by Kestrel Sovereign via the kestrel_sovereign.features entry point — install it alongside kestrel-sovereign and GCPComputeFeature registers itself at startup.

Configuration

Variable	Description
GCP_PROJECT_ID	GCP project ID (required)
GOOGLE_APPLICATION_CREDENTIALS	Path to service-account JSON (or use ADC / Workload Identity)

Optional [gcp_compute] section in kestrel.toml:

[gcp_compute.manager]
default_zone = "us-central1-a"
default_region = "us-central1"
prefer_spot = true

What's provided

• GCPComputeFeature — agent-facing tools to provision/list/stop GPU instances, run training over SSH, monitor jobs
• Standalone API: GCPComputeEngineManager for direct programmatic use
• Spot-instance preference + automatic lifecycle (TTL, budget caps)

Dependencies

• kestrel-sovereign-sdk>=0.2,<1 — base Feature, tool, ToolCategory, BackendType interfaces
• kestrel-sovereign>=0.5,<1 — kestrel.toml unified-config loader (runtime)
• google-cloud-compute>=1.15.0
• google-cloud-logging>=3.5.0

Development

uv pip install -e '.[test]'
uv run pytest

License

Apache-2.0