Vectara MCP Server for Claude

Overview

The Vectara MCP Server is the official Model Context Protocol server from Vectara, providing grounded RAG (Retrieval-Augmented Generation) capabilities to Claude. It exposes six tools: RAG query with answer generation, semantic search over document chunks, hallucination detection, factual consistency evaluation, and dynamic API key management. The corpus key is passed as a tool argument at query time (not an env var), making it easy to query different corpora in the same session. Licensed under Apache-2.0.

Key capabilities

• RAG query — ask natural language questions and get grounded answers with source citations.
• Semantic search — retrieve ranked document chunks without LLM answer generation.
• Hallucination correction — detect and fix hallucinations using Vectara's VHC API.
• Factual consistency — evaluate whether a claim is factually consistent with source docs.
• Session API key management — configure or clear the API key within a session.

Tools

How it compares

Vectara's built-in hallucination correction (VHC API) is unique — no other RAG MCP server offers an integrated hallucination detection and correction tool.

Installation

Claude Code

claude mcp add vectara \
  -e VECTARA_API_KEY=<your-api-key> \
  -- uvx vectara-mcp --stdio

Claude Desktop

{
  "mcpServers": {
    "vectara": {
      "command": "uvx",
      "args": ["vectara-mcp", "--stdio"],
      "env": {
        "VECTARA_API_KEY": "<your-api-key>"
      }
    }
  }
}

The corpus key (VECTARA_CORPUS_KEY) is a tool argument passed at query time — ask Claude to query a specific corpus by name or key rather than setting it as a static env var.

Requirements

• A Vectara account with at least one populated corpus.
• API key with query permissions.
• Python with uvx (from uv).
• An MCP client (Claude Code or Claude Desktop).

Security

• VECTARA_API_KEY grants query access to your corpora — use a scoped key with only read/query permissions.
• Document content retrieved from corpora enters Claude's context — ensure sensitive documents are not in public or shared corpora.

Source Verification Notes

Verified on 2026-06-18:

• Official repository vectara/vectara-mcp (Apache-2.0) on PyPI as vectara-mcp documents the uvx vectara-mcp --stdio install, VECTARA_API_KEY environment variable, corpus key as a tool argument (not an env var), all six tools (ask_vectara, search_vectara, correct_hallucinations, eval_factual_consistency, setup_vectara_api_key, clear_vectara_api_key), and the hallucination correction via Vectara's VHC API.
• Claude Code MCP documentation at code.claude.com/docs/en/mcp describes the stdio connector pattern used above.