MCP servers for a RAG agent
The MCP server bundle to build an AI agent that answers from your own documents with citations.
3 servers · ~22 tools · Last updated June 17, 2026
TL;DR: A retrieval-augmented agent needs three things: a way to gather content, a vector store to remember it, and current docs to stay accurate. This bundle covers all three so the agent answers from your data instead of hallucinating.
Bottom line: start with Chroma MCP Server and add the rest as your needs grow. All 3 install together via the merged config below (~22 tools total).
Tool budget: this stack exposes about 22 tools. That's within Cursor's practical ~40-tool ceiling, so all servers can stay enabled together. Check your own config →
What's in the stack
Vector store for embeddings — the retrieval memory at the core of RAG.
Source ↗Injects version-correct library docs so code answers don't drift.
Source ↗One-click config
All 3 servers merged into a single block — pick your client and paste.
{
"mcpServers": {
"chroma-mcp-server": {
"command": "uvx",
"args": [
"chroma-mcp",
"--client-type",
"persistent",
"--data-dir",
"/absolute/path/to/data"
]
},
"firecrawl": {
"command": "npx",
"args": [
"-y",
"firecrawl-mcp"
],
"env": {
"FIRECRAWL_API_KEY": "<your-firecrawl-api-key>"
}
},
"context7-mcp-server": {
"command": "npx",
"args": [
"-y",
"@upstash/context7-mcp"
],
"env": {
"CONTEXT7_API_KEY": "<optional-your-context7-api-key>"
}
}
}
}Paste into ~/Library/Application Support/Claude/claude_desktop_config.json and fully restart Claude Desktop. Replace any placeholder keys/tokens with your own.
Capabilities this stack covers
FAQ
What's the minimum stack for RAG?
A vector store plus a way to ingest content. Add a docs server if the agent writes code, and a web-search server if it needs live information.
Can I swap the vector store?
Yes — Chroma, Qdrant or Pinecone all fit the same slot. Pick on scale and hosting preference.