MCP 服务器

knowledgeplane

MCP server that gives AI agents and teams persistent, shared memory using a knowledge graph with vector embeddings, automatic consolidation of related facts, and hybrid search.

README

<h1 align="center">Knowledge Plane</h1>

<p align="center"> Shared memory for AI agents and teams -- stops your tools from forgetting. </p>

What is Knowledge Plane?

Knowledge Plane is an MCP server that gives AI agents and teams persistent, shared memory. It stores facts as a knowledge graph with vector embeddings, automatically consolidates related facts into knowledge cards, and provides hybrid search (vector + graph traversal). Native MCP protocol support means any MCP-compatible AI tool (Claude, Cursor, etc.) can read and write shared knowledge.

Why Knowledge Plane?

Memory layers for AI exist (Mem0, Zep, Letta), but they're built around flat vector stores -- good for retrieval, less good for understanding how things relate. Knowledge Plane stores facts as a graph with typed relations, so "Alice manages Bob" is a first-class edge you can traverse, not a blob of text you hope cosine similarity resurfaces. Background workers continuously consolidate related facts into knowledge cards, building institutional knowledge instead of just accumulating embeddings. And it's designed for teams from the start: multi-workspace, multi-agent, multi-tool -- shared memory across your entire AI stack, not just one chatbot.

How It Works

Your team runs AI agents across different tools -- a dev agent in the IDE, a PM agent in the browser, a support agent in Slack. On Monday, the dev agent learns that the payments API was refactored. On Wednesday, the PM agent is planning a sprint and needs that context. Without shared memory, it doesn't have it. With Knowledge Plane, the dev agent wrote that fact once, the consolidation worker linked it to the payments service card, and now any agent in any tool can find it through search or graph traversal.

Open Infrastructure

Built on MCP (Model Context Protocol), the emerging open standard for AI tool integration. Fully self-hostable -- your knowledge stays on your infrastructure. Apache-2.0 licensed, no telemetry, no vendor lock-in. Your data, your rules.

Built by Camplight, a cooperative of AI-native teams across Europe. Used internally to power knowledge sharing across Juma.ai and Slack-integrated marketing agents.

Features

Persistent agent memory -- Facts stored in ArangoDB knowledge graph with vector embeddings
MCP-native -- Drop-in memory for Claude Desktop, Cursor, and any MCP client
Hybrid search -- Vector similarity + graph traversal + optional BM25
Auto-consolidation -- Background workers merge related facts into knowledge cards
File upload and extraction -- Upload documents, extract facts automatically
Web dashboard -- Browse, search, and manage your knowledge graph
Multi-workspace -- Isolated knowledge spaces per team or project
OAuth + API keys -- Google/GitHub OAuth and API key authentication
REST API -- Full CRUD API alongside the MCP server
Audit trails -- Track who created and modified every fact

Quick Start

# Clone and install
git clone https://github.com/camplight/knowledgeplane.git
cd knowledgeplane
npm run bootstrap

# Set up environment
./scripts/setup-env.sh    # Creates .env files from examples

# Start everything (ArangoDB + all services)
npm run dev

Services start at:

Service	URL
Web Dashboard	http://localhost:3000
MCP Server	http://localhost:8080
REST API	http://localhost:8081
ArangoDB	http://localhost:8529

For detailed setup including OAuth and ngrok, see Getting Started.

MCP Integration

Claude Desktop

Add to your Claude Desktop MCP config:

Option 1: HTTP/SSE (recommended)

{
  "mcpServers": {
    "knowledgeplane": {
      "url": "http://localhost:8080/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_API_KEY"
      }
    }
  }
}

Option 2: stdio adapter

{
  "mcpServers": {
    "knowledgeplane": {
      "command": "node",
      "args": ["apps/mcp-server/dist/mcp/adapter.js"],
      "env": {
        "KNOWLEDGEPLANE_API_URL": "http://localhost:8080",
        "KNOWLEDGEPLANE_API_KEY": "YOUR_API_KEY"
      }
    }
  }
}

Architecture

knowledgeplane/
├── apps/
│   ├── mcp-server/          # MCP protocol server (Fastify)
│   ├── rest-api/            # REST API (Express)
│   ├── webapp/              # Web dashboard (Next.js)
│   └── background-workers/  # Consolidation & embeddings
├── packages/
│   ├── db/                  # ArangoDB models & queries
│   ├── aimodel/             # LLM abstraction layer
│   ├── api-core/            # Shared API utilities
│   └── file-processor/      # Document parsing & extraction
└── infra/                   # Docker Compose configs

Documentation

Guide	Description
Getting Started	Quick start and setup
Development Guide	Local development with ngrok
Deployment Guide	Cloud deployment (Railway, etc.)
Environment Setup	All environment variables
API Specification	Complete API reference

Knowledge Plane Cloud

This repo is the self-hosted, DIY version. If you'd rather not manage the infrastructure yourself, Knowledge Plane Cloud offers a managed deployment with everything pre-configured. And if you need a full agentic memory overhaul -- integrating shared knowledge across your team's AI tools and workflows -- Camplight (the team behind Knowledge Plane) does that as a service.