Vectara MCP: connect agents to Trusted RAG

Who it is for

If you want an agent app connected to Vectara Trusted RAG—fast, stable retrieval-augmented answers with fewer hallucinations—this service matches. Official docs describe that capability over the MCP protocol.

It fits people who can already pip-install a package and hold a Vectara API key. Default startup is authenticated HTTP transport. Without a key, the session-level configuration tool cannot pass validation.

The first time, only validate the key. Do not treat query tools that have not been named as known functions. Retrieval still goes online and also sends your question to Vectara.

First try in 3 minutes

The official recommended path is to install from PyPI and start with the default secure HTTP transport. Prerequisite: you already have a Vectara API key.

Install:

pip install vectara-mcp

Set the environment variable VECTARA_API_KEY to the real key issued by the console. Official sample values are placeholders; do not copy them. Then start:

python -m vectara_mcp

Official docs say the service by default runs at:

http://127.0.0.1:8000

and authentication is already on. Point the MCP client at that HTTP service. Official docs do not write the full header fields to put in the client; do not invent them. Connect the way the client connects to authenticated HTTP MCP.

Then make one call: call setup_vectara_api_key. Official docs say it configures and validates the Vectara API key for the session; it is a one-time setup. Parameter fields are not written; fill in the real key according to the definition the client shows, and do not write sample strings.

Success signal: validation passes and the session can use that key. If it fails, first check that the environment variable and the tool argument are the real key.

Do not add --stdio. The first time, stay on default HTTP.

Usage examples

Example 1: validate the key inside the session. Call setup_vectara_api_key. The official note is Configure and validate your Vectara API key for the session. It does not write the key format or whether it overrides the environment variable. On success you should see validation pass. Do not treat it as a tool that can query a corpus.

Example 2: after validation, run a Trusted RAG query. Official docs say this server gives agent apps Trusted RAG. This page only names the setup tool above and does not write a query function name; open the tools this server shows in the client and pick the item described as retrieval or Q&A. Do not invent function names that never appeared. On success you should see retrieval results from Vectara. Queries go online and may also expose your question text.

Limits

Default is authenticated HTTP. There is also python -m vectara_mcp --stdio, which official docs mark as less safe, only for local development or Claude Desktop, not the first-try path.

VECTARA_API_KEY is required. Official docs write neither a rate limit nor undo. This page does not name query tools; do not treat the words Trusted RAG as a function name. The local loopback address only means the default listen address; it does not mean data stays on your machine—retrieval still hits Vectara.