Skip to main content
MCP server CX Codex

veritasgraph — Claude Code MCP server: GraphRAG with citations

An on-prem GraphRAG server that answers from a knowledge graph with source citations and admits when evidence is missing.

Install into Claude Code · Platform CLI All methods and platforms ↓
claude mcp add --env VERITASGRAPH_MODEL=<VERITASGRAPH_MODEL> --env OLLAMA_HOST=<OLLAMA_HOST> --env STUDIO_DATA_DIR=<STUDIO_DATA_DIR> --transport stdio veritasgraph -- uvx veritasgraph-mcp
Open on GitHub

What veritasgraph does

VeritasGraph answers by traversing a knowledge graph instead of ranking chunks by similarity. Provenance is stored when documents are written, so citations are not invented at generation time. If the graph has no edge for a question, the agent returns insufficient evidence rather than a guess. Traversal is reproducible, which keeps the evidence trace separate from the model's own output.

It runs as an MCP server through uvx veritasgraph-mcp. Ingest and queries need a local Ollama. Configuration takes VERITASGRAPH_MODEL, OLLAMA_HOST and STUDIO_DATA_DIR. The demo modes differ in what they ask of you: --mode=lite works with an OpenAI or Anthropic API key, --mode=local wants Ollama and 8GB RAM, and --mode=full needs Docker and Neo4j.

There is a worked example, MedTrust Agent, that applies the same architecture to medical and pharmaceutical workflows. It requires [doc#chunk] citations and withholds uncited answers, keeps a tamper-evident audit log, and blocks any tool call that fails authorization. Worth a look if you need that kind of control plane around agent tools.

What it helps with

  • get answers tied to specific document fragments
  • keep a reproducible evidence trace apart from model generation
  • make the agent admit when the graph lacks evidence
  • run graph-based retrieval entirely on your own hardware
  • audit which tools an agent actually invoked

Example prompts for the agent

  • add veritasgraph as an MCP server and check that it responds
  • point veritasgraph at local Ollama and set the model via VERITASGRAPH_MODEL
  • ask the graph which documents support this conclusion

How to install veritasgraph

Platform CLI

claude mcp add --env VERITASGRAPH_MODEL=<VERITASGRAPH_MODEL> --env OLLAMA_HOST=<OLLAMA_HOST> --env STUDIO_DATA_DIR=<STUDIO_DATA_DIR> --transport stdio veritasgraph -- uvx veritasgraph-mcp

Replace <…> with your own values. uvx runs the command named after the package; if the README names another one, use uvx --from veritasgraph-mcp <command>

Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more

Questions about veritasgraph

How do I install veritasgraph in Claude?

Run claude mcp add veritasgraph -- uvx veritasgraph-mcp. For environment variables use the longer form: `claude mcp add --env VERITASGRAPH_MODEL=<VERITASGRAPH_MODEL> --env OLLAMA_HOST=<OLLAMA_HOST> --env STUDIO_DATA_DIR=<STUDIO_DATA_DIR> --transport stdio veritasgraph -- uvx veritasgraph-mcp`.

What does veritasgraph do?

It is an MCP server for GraphRAG: retrieval walks a knowledge graph rather than comparing chunk vectors, and each answer carries citations. When the graph lacks the needed edges, it says so instead of guessing.

What do I need to run veritasgraph?

A local Ollama handles ingest and queries. The --mode=lite demo only needs an OpenAI or Anthropic API key, --mode=local requires Ollama and 8GB RAM, and --mode=full needs Docker plus Neo4j.

Similar MCP servers

All MCP servers for AI and ML →

Project sponsors

Companies that keep this analytics open
Gold sponsor

Impressions, clicks and code copies — open statistics. We keep the numbers public for the same reason NeBlask itself is public. Become a sponsor