Compute
Governed compute for unified-memory AI hardware. Nothing starts unless it provably fits, and a watchdog kills the newest model before the box wedges.
Admission control against a hard memory budget, Scenes that activate a whole model set transactionally, a 1 Hz watchdog on MemAvailable, a CLI, and an MCP server so agents drive the node as tools. Public under Apache-2.0 at git.karti.ai/lumbridge-public/compute.