Grimoire

Grimoire is a skill discovery server that exposes reusable prompt workflows as MCP tools. LLM agents can search for relevant skills, retrieve step-by-step instructions, and then execute the suggested tools via MCPBlade or other MCP servers.

How It Works

LangChain Agent
    │
    ├── grimoire (knowledge layer)
    │   └── search_skills("deploy to k8s") → skill with prompt + suggested tools
    │
    └── mcpblade (execution layer)
        └── call suggested tools (kubectl_apply, docker_build, ...)

A Skill is a structured prompt template that tells the LLM what to do and which tools to use:

id: deploy_k8s
name: Deploy to Kubernetes
description: Guide the deployment process to a Kubernetes cluster
tags: [kubernetes, deploy, k8s, container]
suggestedTools:
  - kubectl_apply
  - kubectl_get
  - docker_build
prompt: |
  Please follow these steps to deploy to Kubernetes:
  1. Use docker_build to build the container image
  2. Use kubectl_apply to apply the deployment manifest
  3. Use kubectl_get to verify the deployment status

MCP Tools

Grimoire exposes three tools via the MCP protocol:

Tool	Description
search_skills	Semantic search for skills using natural language queries
find_skill	Find a specific skill by name, returns full prompt and suggested tools
list_skills	List all available skills

Installation

go install github.com/flarexio/grimoire/cmd/grimoire@latest

Configuration

Create ~/.flarex/grimoire/config.yaml:

skillsDir: skills
vector:
  enabled: true
  persistent: true
  collection: skills

Place skill definitions as YAML files in the skills/ directory.

Usage

# Run with default configuration (~/.flarex/grimoire)
grimoire

# Run with custom path
grimoire --path /path/to/config

# Custom HTTP address
grimoire --http-addr :9090

Integration with MCPBlade

Register Grimoire as a backend MCP server in MCPBlade's config.yaml:

mcpServers:
  grimoire:
    transport: streamable-http
    url: http://localhost:8080/mcp/

MCPBlade aggregates Grimoire's tools alongside other MCP servers, providing a unified interface for LLM agents.

Architecture

grimoire/
├── skill/                  # Domain model
│   ├── skill.go            # Skill struct and errors
│   └── repository.go       # Repository interface
├── service.go              # Core business logic
├── endpoint.go             # Go-Kit endpoints
├── logging.go              # Logging middleware
├── model.go                # Config and vector document helpers
├── mcp/
│   └── endpoint.go         # MCP protocol endpoints
├── transport/http/         # HTTP transport layer
├── persistence/
│   ├── chromem/            # Vector DB (ChromeM) implementation
│   └── yaml/              # YAML file-based skill repository
└── cmd/grimoire/           # Entry point

API Endpoints

GET  /api/skills            # List all skills
GET  /api/skills/search     # Search skills (?query=...&k=...)
GET  /api/skills/:name      # Find skill by name
POST /mcp/                  # MCP streamable HTTP endpoint

Dependencies

• mcp-go — MCP protocol
• chromem-go — Vector database
• go-kit — Microservice toolkit
• gin — HTTP framework
• zap — Structured logging

License

MIT License — see LICENSE.md for details.

This project is part of the FlareX ecosystem.