garnet-mcp server

<!-- mcp-name: io.github.tarazou9/garnet-mcp -->

Garnet Vector Memory MCP Server

An MCP (Model Context Protocol) server that gives LLM agents a retrieval-augmented (RAG) memory backed by Garnet Vector Sets. Agents store text as embeddings and later retrieve the most semantically relevant snippets by meaning to ground their responses — persisting across sessions, not just within a single conversation. This is application-layer semantic recall (RAG): it augments the model's prompt with retrieved context, and is distinct from inference-layer KV-cache reuse (e.g. LMCache). It runs against a local or self-hosted OSS Garnet (or any Redis-compatible endpoint that supports Vector Sets), with bring-your-own embeddings.

Tools

The server exposes five tools over the MCP stdio transport (HTTP is also supported):

Prerequisites

• .NET 10 SDK — required so clients can launch the server with the dnx command (ships with the .NET 10 SDK). The package itself targets net9.0.
• A Garnet server started with Vector Sets enabled. Vector Sets are currently a Garnet preview feature and must be turned on explicitly with --enable-vector-set-preview. Any Redis-compatible endpoint exposing the V* vector commands also works.
• Embeddings (optional but recommended). Set EmbeddingsProvider=AzureOpenAI and point it at an Azure OpenAI embedding deployment (keyless auth via DefaultAzureCredential / az login). Without it, the server falls back to a deterministic Fake** provider that requires no external calls — useful for local demos and tests, but not for meaningful semantic recall.

Install (from NuGet)

The server is published as a .NET tool package that MCP clients run via dnx. Add it to your client's MCP config, e.g. .vscode/mcp.json for VS Code / GitHub Copilot:

{
  "servers": {
    "garnet-mcp": {
      "type": "stdio",
      "command": "dnx",
      "args": ["GarnetMcp.Server@0.1.0", "--yes"],
      "env": {
        "Garnet__Host": "127.0.0.1",
        "Garnet__Port": "6379",
        "Embeddings__Provider": "AzureOpenAI",
        "Embeddings__Endpoint": "https://YOUR-RESOURCE.openai.azure.com/",
        "Embeddings__DeploymentName": "text-embedding-3-small",
        "Embeddings__Dimensions": "1536"
      }
    }
  }
}

To run the fully offline demo, drop the Embeddings entries (or set EmbeddingsProvider=Fake).

Recommended agent instructions

The server already sends usage guidance to any MCP client at connect time (via the server's ServerInstructions), so most agents will use the tools appropriately on their own. If you want to make an agent use the memory proactively, add something like the following to that agent's custom-instructions (e.g. .github/copilot-instructions.md in your project, a Claude/Cursor rules file, etc.):

Before answering anything about the user — their preferences, facts, decisions, or past statements — FIRST call recallmemory, even if the answer seems to be in the current chat (memories persist across sessions). Whenever the user states a durable fact about themselves or asks you to remember something, call storememory to persist it, then briefly confirm what you saved.

Configuration

All settings are read from configuration / environment variables. In environment-variable form, use __ (double underscore) as the section separator.

Shortened. The full README is on GitHub.