Kappa Graph MCP
A knowledge graph with an MCP server: the agent searches concepts by meaning and sees how well each is supported.
- 131 ★ on the repository
- Last commit
Extended access
Why this level
- The `graph`, `ingest` and `ontology` tools can add, change and delete platform data.
- Document text goes to the chosen language model during concept extraction.
Install
- Works in
Manual install
curl -fsSL https://raw.githubusercontent.com/aaronsb/knowledge-graph-system/main/client-manager.sh | bashThe first README method installs the `kg` CLI, the MCP server and FUSE; CLI only: `npm install -g @aaronsb/kg-cli`. The platform is deployed separately through `./operator.sh`.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The `graph`, `ingest` and `ontology` tools can add, change and delete platform data.
- Document text goes to the chosen language model during concept extraction.
This is third-party code. Review the repository files before installing.
What it does
Kappa Graph (the `knowledge-graph-system` repository) extracts concepts, relations and evidence from documents, stores them in a graph on PostgreSQL with Apache AGE and computes a grounding score from -1 to 1 for each concept. Contradicting sources are kept side by side rather than merged. The `kg-mcp-server` MCP server gives an agent the tools `search`, `concept`, `ontology`, `job`, `ingest`, `source`, `graph`, `program`, `document` and more: search by meaning, relation traversal, paths between concepts, text ingestion and job management. The platform is deployed on your own machine with `./operator.sh init` and `./operator.sh start`. Client tools (the `kg` CLI, the MCP server and a FUSE filesystem) are installed by the `client-manager.sh` script.
Who it is for. Researchers and engineers who accumulate documents and want to give an agent memory with a reliability score and a source link.
Good fit when
- An agent must answer from a large set of papers and show what a conclusion rests on.
- Places where sources disagree must be found in notes and reports.
- The shortest chain of relations between two concepts is needed.
- Texts and images from a project are loaded into a shared ontology.
Not a fit when
- A single MCP server without deploying a platform and containers is needed.
- Documents must not go to a cloud language model and local Ollama is unavailable.
- An agent must not be trusted to change and delete graph concepts.
Example request
Which concepts relate to distributed systems and how well are they supported? Find the shortest path from technical debt to team productivity.MCP
- Transport
- stdio
- Authentication
- OAuth
Environment variables
KG_API_URLrequired- Platform API address, for example `http://localhost:8000`.
KG_OAUTH_CLIENT_IDrequired- ID of the OAuth client created by `kg oauth create-mcp`.
KG_OAUTH_CLIENT_SECRETrequired, secret- OAuth client secret, shown once at creation.
Limitations
A running platform is required: containers, PostgreSQL and your own installation, not just the MCP server. Concept extraction uses a language model: a cloud provider receives document text, while the local option runs through Ollama. Node.js 20.12.0 or newer and an OAuth client created with `kg oauth create-mcp` are required; the secret is shown once. The MCP server is installed together with the `kg` CLI, so both the client tools and a running platform backend are needed.
How to disable. Remove the server with `claude mcp remove knowledge-graph` or by editing `claude_desktop_config.json` and delete the OAuth client from the platform. `./operator.sh stop` stops the platform itself.
FAQ
What is the grounding score?
It is a number from -1 to 1: supporting evidence raises it and contradicting evidence lowers it. A concept with one unchallenged mention scores differently from one with dozens of sources.
How do I create MCP credentials?
Run `kg login`, then `kg oauth create-mcp`. The command prints `KG_OAUTH_CLIENT_ID` and `KG_OAUTH_CLIENT_SECRET`; the secret is visible only once. The server refreshes the one-hour tokens itself.
Can it work without cloud models?
Yes, the platform supports local inference through Ollama, so a deployment can run without external services.
Which files can be ingested?
PDF, Markdown, images and plain text. Concepts, relations and evidence are extracted from them, and each concept stays linked to its source fragment.
What is the FUSE filesystem?
An optional client that mounts the graph as a directory: `ls`, `grep` and `find` work over semantic space. The same `client-manager.sh` script installs it.
Discussion
No questions or reviews yet.
Related
MCP serversMCP server for the Flexible GraphRAG platform: ingest from 14 data sources, hybrid search and AI answers over a knowledge graph
A CLI and MCP server that extracts structured knowledge from documents: graphs, hypergraphs, and fact search with sources
The official Graphlit MCP server: an agent collects material from Slack, GitHub, mail and websites into a knowledge base and searches it
InfraNodus MCP server turns any text into a knowledge graph and shows topic clusters, the gaps between them and research questions
A local MCP server that turns a folder of documents into a knowledge base with hybrid search for Claude Code and Cursor
A graph database with git-style branches: the server exposes it to agents over MCP at /mcp, and a skill teaches the agent schemas and queries
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Kappa Graph (this card) | no | OAuth | 131 | |
| Cognee | no | 33,093 | a knowledge graph for agent memory; in Kappa Graph every concept gets a grounding score and contradictions between sources are kept | |
| Graphiti MCP | no | API key | 32,574 | a temporal knowledge graph over MCP; Kappa Graph builds an ontology from documents and measures concept grounding |
README in short
The README covers a quick start for client tools and the platform, an architecture on PostgreSQL with Apache AGE and the `graph_accel` extension, a comparison with GraphRAG, Zep/Graphiti and Cognee, and documentation of 96 architecture decision records.