Pinecone MCP Server

Pinecone MCP Server is Pinecone's official Model Context Protocol integration. It connects MCP-capable coding assistants such as Cursor, Claude, and Claude Code directly to Pinecone projects and their documentation. According to the official documentation, an agent can use it to search Pinecone documentation, list and describe indexes, create new indexes, upsert records, search semantically, and merge and rerank results across multiple indexes.

Purpose: developer workflow with a vector database

According to official documentation, Pinecone MCP Server targets developers who use Pinecone as part of their technology stack. It is designed for use in coding assistants, not as a general-purpose AI agent. Its focus covers three areas. First, searching the official Pinecone documentation so that the assistant can answer questions about the API, index configuration, and best practices without relying on pre-trained knowledge. Second, managing indexes, including creation, description, and status queries. Third, working with data: upserting records and running searches directly from the development environment to test queries and evaluate results.

Tools: what the assistant can execute

The official documentation describes nine tools provided by the MCP server. search-docs searches the official Pinecone documentation. list-indexes returns all indexes in the connected Pinecone project. describe-index describes the configuration of a single index. describe-index-stats provides statistics about the contents of an index, including the record count and available namespaces. create-index-for-model creates a new index that uses an integrated inference model to embed text as vectors. upsert-records inserts or updates records in an index with integrated inference. search-records searches for records based on a text query with embedding and supports metadata filtering and reranking. cascading-search searches across multiple indexes, deduplicates, and reranks the results. rerank-documents reorders a collection of records or text documents using a specialized reranking model.

According to the documentation, the MCP server supports only indexes with integrated inference. Indexes without integrated inference, standalone embeddings, and plain vector search are not supported.

Prerequisites and setup

Connecting to a Pinecone project requires a Pinecone API key, which can be created through the Pinecone console. Without an API key, the assistant can still search documentation but cannot manage or query indexes. For local execution, the server requires Node.js v20 or later, with node and npx available in PATH. The npm package is @pinecone-database/mcp, at version 0.3.0 according to the npm registry.

Configuration happens through a JSON file specific to each client. For Cursor, create a .cursor/mcp.json file in the project root. In Claude Desktop, add the settings under Settings > Developer > Edit Config. Claude Code uses the claude mcp add-json command. Antigravity provides an MCP Store for direct installation. For Gemini CLI, the repository documents an extension install command. The PINECONE_API_KEY is placed in the env section of the server configuration in all cases.

Boundary: MCP server versus Assistant MCP

Official Pinecone documentation distinguishes the Pinecone Developer MCP Server from a separate Assistant MCP Server. The Developer MCP Server improves the workflow of developers who use Pinecone in their stack. The Assistant MCP Server gives AI assistants access to context sourced from uploaded files in a Pinecone knowledge base. Both products are independent offerings with different use cases and configuration paths.

Security and data handling

The PINECONE_API_KEY should be stored exclusively in the client's secret or credential management facility, never in a source code repository, shared configuration files, prompts, or in plaintext in terminal history or screenshots. Tool calls that execute index management or write operations make permanent changes to the connected Pinecone project. This includes creating indexes and upserting records. For write operations, obtain approval from a responsible person before the agent acts. Restrict agent permissions to the operations that are actually needed.

Records upserted through upsert-records leave the local device and are stored on Pinecone's servers. The same applies to text queries sent to search-records, which are processed by Pinecone's integrated inference to produce embeddings. Before using sensitive data, review Pinecone's current terms of service and privacy policy directly with the provider. Both the connected MCP client and the assistant's model provider may additionally process and retain prompt content and tool results.

The official pinecone-io/pinecone-mcp repository is Apache-2.0 licensed. On 2026-09-08, the GitHub API reported exactly 71 stars. Stars are a point-in-time repository-popularity signal only; they do not demonstrate source quality, privacy, completeness, or security.

FAQ

Can the server also access Pinecone Assistants and their knowledge base? No. According to official documentation, there is a separate Pinecone Assistant MCP Server for that purpose. The Pinecone Developer MCP Server is designed for working with indexes and developer documentation.

Does the server work without an API key? In a limited way. Without a key, the assistant can search Pinecone documentation according to the documentation, but it cannot perform any index operations.

Which indexes are supported? Only indexes with integrated inference, according to the documentation. Indexes using external embedding models are not supported by this MCP.

An MCP-capable coding client, Node.js v20+ with npx, and a Pinecone API key for index operations.

Set PINECONE_API_KEY as a secret. For Cursor: create .cursor/mcp.json with the npx command. For Claude Desktop: Settings > Developer > Edit Config. For Claude Code: use claude mcp add-json. For Antigravity: install from the MCP Store.

npx -y @pinecone-database/mcp

Pinecone API key, stored as PINECONE_API_KEY in the env section of the client configuration. Without a key, only documentation search is available.

Access and available operations depend on the Pinecone account, API key permissions, client support, and current provider terms.

Tool calls use the Pinecone API; upserted records and search queries are processed on Pinecone servers. Prompts and results can also reach the model and logging path of the connected client.

API keys can be exposed through configuration files, logs, or repositories. Write operations permanently modify the Pinecone project and require human approval. Sensitive data should not be upserted without reviewing provider terms.