Happy slice of mcp pizzaMCP.pizza

mcp-graphiti

MCP.Pizza Chef: rawr-ai

Notes you hand over are read by a language model, broken into people, facts, and connections, then stored in a database you run yourself so later questions can be answered from them. Budget for it: an OpenAI key with credit on it is required, and every note you file costs a paid call. Eight commands cover adding, searching, fetching, and deleting. Wiping everything needs a two-step confirmation that really is enforced, though single deletions are not protected at all. The code has not changed since August 2025.

Unmaintained · No commits in 12 months.
Data
Notes

Use This MCP server To

Save meeting notes so I can ask about them later Ask who was involved in a decision months ago Search past notes for what I know about a client Keep separate memories for separate projects Remove a note that turned out to be wrong

README

Graphiti MCP Server

Fork of the getzep/graphiti example with a focus on developer experience and multi‑project support. Graphiti extracts entities and relationships from text and stores them in Neo4j. This repo adds a CLI that spins up a root server plus project‑specific MCP servers in Docker so several knowledge graphs share the same database.

Quick Start

  1. Install and clone 
    pipx install 'git+https://github.com/rawr-ai/mcp-graphiti.git'
git clone https://github.com/rawr-ai/mcp-graphiti.git
cd mcp-graphiti
cp .env.example .env  # fill in Neo4j credentials and your OpenAI key
  2. Launch services 
    graphiti compose   # generates docker-compose.yml and updates .cursor/mcp.json
graphiti up -d
    The root server runs on port 8000; project containers start at 8001.
  3. Create a project 
    cd /path/to/my-kg
graphiti init my-kg        # writes ai/graph/mcp-config.yaml
# add entity definitions under ai/graph/entities/
    Rerun graphiti compose && graphiti up -d from anywhere to start its container.

Once running you can:

  • Check http://localhost:8000/graphiti/status.
  • Connect MCP‑compatible tools to http://localhost:800{N}/sse.
  • Browse Neo4j at http://localhost:7474 using the credentials in .env.

Security note

If NEO4J_PASSWORD remains password the server refuses to start unless GRAPHITI_ENV=dev. Always use a strong password in production.

Why this fork?

The upstream repository assumes one server per compose file. Here a single compose file manages many project servers that share Neo4j. Each service gets its own group_id, entities and model so projects stay isolated while running on the same database.

Highlights

  • Project isolation – different extraction rules or models never collide.
  • Editor auto‑discovery – ports are written to .cursor/mcp.json.
  • Crash containment – a bad prompt only restarts its container.
  • Hot reload – tweak a project's config and run graphiti reload <container>.

Leave mcp-projects.yaml empty if you only need the root server.

Danger zone

Setting NEO4J_DESTROY_ENTIRE_GRAPH=true wipes all projects the next time you run graphiti up. Use with care.

Contributing

PRs and issues are welcome.

© 2025 rawr‑ai • MIT License

mcp-graphiti FAQ

Do I need a paid account?
Yes. An OpenAI key with credit is required, because every note you save is put through a paid model to pull out the names and connections.
What will it cost me?
The database runs on your own machine for free, but the per-note charges add up quickly if you file a lot of text, so watch your OpenAI balance.
Which apps does it work in?
Cursor is wired up for you automatically, and any app that can connect to a local web address will work.
Can I use this to remember what a client told me?
Yes — paste the notes in once, then later ask questions about people, dates, and decisions instead of re-reading everything.
Can it delete my notes without asking?
Wiping the whole memory is protected by a two-step confirmation the code genuinely enforces, which is unusual and good. Removing a single note or fact has no such guard and happens immediately.
Is anything exposed to my network?
Yes. The port is published to every device on your network with no password on it, so avoid running this on shared or public Wi-Fi.
How hard is setup?
Genuinely developer-level. You need Docker, a Neo4j database, an environment file, and a command-line tool that generates the configuration.
Is there any way to lose everything at once?
Yes. A setting named NEO4J_DESTROY_ENTIRE_GRAPH erases every project's data the next time you start it, so leave it switched off.
Is it maintained?
The last code change was August 2025. It is a fork of the example server that ships with the Graphiti project.