Skip to main content
Glama

๐ŸŒŸ Why VramArmor?

Running open-source models like DeepSeek-R1, Llama 3.3, and Qwen 2.5 locally is incredibleโ€”until you hit CUDA: Out of Memory (OOM) errors, or your model silently spills into CPU system RAM, dropping token generation speeds from 50 t/s down to an unusable 2 t/s.

VramArmor is a zero-runtime-dependency CLI and Node 24 tool that gives you complete visibility into your GPU VRAM, predicts whether a model will fit before you download or load it, and lets AI agents optimize local inference memory via Model Context Protocol (MCP).

  • โšก Zero External Dependencies: Built with native Node 24 ESM and system GPU probes (nvidia-smi, Apple Silicon unified memory).

  • ๐Ÿงฎ Pre-Flight VRAM Calculator: Calculates model weights, KV cache buffers across context sizes (2k to 128k), and CUDA runtime overhead to prevent OOM panics.

  • ๐Ÿ›‘ Instant Model Memory Reclaimer: Directly inspects and offloads idle models from Ollama / vLLM to free VRAM immediately.

  • ๐Ÿ”ฎ Native Model Context Protocol (MCP) Server: Equips Claude Desktop and Cursor to inspect GPU load and choose models that fit your hardware.

  • ๐Ÿš€ Enterprise Multi-GPU Shard Balancer (BSL 1.1): Calculates proportional layer offloading across multi-GPU rigs (dual RTX 4090s, A100 clusters).


Related MCP server: vram-mcp

๐Ÿš€ Quick Start

Installation

# Run on-demand without installing
npx -y vramarmor help

# Or install globally
npm install -g vramarmor

๐Ÿ’ป CLI Commands

1. Inspect Live GPU VRAM & Hardware Status

vramarmor status

2. Predict VRAM Fit Before Loading

# Check if DeepSeek-R1 14B fits with 8k context in 4-bit
vramarmor predict deepseek-r1:14b Q4_K_M 8192

# Check if Llama 3 70B fits in your card
vramarmor predict llama3:70b Q4_K_M 4096

3. List Active Models Loaded in VRAM

vramarmor models

4. Offload Idle Models to Free VRAM

vramarmor unload deepseek-r1:8b

๐Ÿ”ฎ Model Context Protocol (MCP) Integration

VramArmor provides a native MCP stdio server. Add it to your claude_desktop_config.json or Cursor MCP settings:

{
  "mcpServers": {
    "vramarmor": {
      "command": "npx",
      "args": ["-y", "vramarmor", "mcp"]
    }
  }
}

Now Claude Desktop and Cursor can directly execute:

  • vramarmor_gpu_status: Returns live VRAM metrics, temperature, and utilization.

  • vramarmor_predict_fit: Checks if an LLM will fit safely in GPU memory.

  • vramarmor_list_models: Lists models currently in VRAM.

  • vramarmor_free_vram: Offloads models to reclaim GPU memory.


๐Ÿ’ผ Commercial & Enterprise Support

VramArmor follows an Open-Core model:

  • Community Core (Apache 2.0): 100% Free for single-GPU developers, VRAM predictor, Ollama manager, and full MCP server.

  • Enterprise Edition (BSL 1.1): Includes the Multi-GPU Balancer (ee/cluster/gpuBalancer.js) for multi-GPU rigs and enterprise clusters with priority SLA.

Direct Contact & Payment


๐ŸŒ The 1nsharma Ecosystem

  • HiveMesh โ€” Autonomous Multi-Agent Workspace Orchestration Mesh & Event Blackboard.

  • DevDaemon โ€” Autonomous Terminal Error Interceptor & Workspace Healer.

  • DroidArmor โ€” Autonomous AI Mobile Agent & Fast ADB Automation for Android.

  • SlashGraph โ€” Universal Codebase & System Architecture Intelligence Graph.

  • TokenArmor โ€” Open-Source AI Gateway & Token FinOps Firewall.

  • 1nsharma/SLASH โ€” Unified System & Digital Network Intelligence Graph.

  • 1nsharma/Urlmind โ€” Understand Anything on the Internet from a Single URL.


Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelogโ€ฆ

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables MCP clients to manage LLM inference memory through virtual paged-attention KV-cache block mapping, non-contiguous physical page allocation, zero-copy fragmentation tracking, radix-trie prefix caching, INT8 quantized compute, and speculative decoding verification. It also exposes prefill and decode latency telemetry so edge deployments can be benchmarked and tuned without external dependencies.
      7
      MIT
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables users to describe an LLM fine-tuning job once and receive ranked GPU recommendations across a dozen cloud providers, calculated from the workload's required VRAM and FLOPs. It surfaces the cheapest, fastest, and most balanced hardware options with clear data provenance, callable from Claude Code, Claude Desktop, a CLI, or a REST API.
      -