Resource Sentinel MCP

Resource Sentinel MCP addresses the challenge of managing resource contention when running multiple coding agents, tests, and models on a single Linux machine. It is designed to prevent common issues such as out-of-memory kills, disk thrashing, and system hangs that occur when workloads oversubscribe available memory and CPU resources. The tool introduces a deterministic control plane, avoiding reliance on heuristic or model-based predictions.

The core functionality combines telemetry collection with lease-based admission control. Resource Sentinel MCP continuously observes the Linux host using psutil-based telemetry, tracking system resource usage in real time. Workloads can request execution slots by submitting their workload ID along with estimated memory and CPU requirements. If the host has sufficient resources, the tool grants an expiring lease for execution; if not, it queues the request until resources become available. All tickets and leases are persisted using SQLite in write-ahead log (WAL) mode, ensuring reliable state tracking.

For integration and monitoring, Resource Sentinel MCP provides FastMCP tools for agents to interact with the control plane and exposes a read-only HTTP projection suitable for dashboards. Additionally, it feeds telemetry compatible with Glitch, enabling visualization or reactive components—such as a "cyber pet"—to reflect the real-time state of the host. This setup is particularly useful for users running parallel workloads like pytest runs, Docker builds, or multiple subagents on shared laptops or lab servers, where resource contention is a frequent problem.

The tool is intended for environments where multiple heavy local jobs are executed concurrently and where deterministic admission control is needed to maintain system stability. It operates on Linux systems and is built to be integrated by agents or workloads that need to manage their resource usage proactively.