# vector-mcp [Health: Active]

**Category:** 🗄️ Databases  
**Repository:** https://github.com/Knuckles-Team/vector-mcp  
**GitHub Stars:** 16  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/vector-mcp

## Description
Vector MCP Server for AI Agents - Supports ChromaDB, Couchbase, MongoDB, Qdrant, and PGVector

## Tools
Capabilities this server exposes over MCP:

- **vector_collection_management** — Manage collection management operations.
- **vector_search** — Manage search operations.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "vector-mcp": {
    "command": "uvx",
    "args": ["--from","vector-mcp[mcp]","vector-mcp"]
  }
}
```

## Documentation

## What vector-mcp does

vector-mcp provides an MCP interface for vector collection management and retrieval. Its default backend is an epistemic-graph engine, while optional provider packages add PostgreSQL/pgvector, Qdrant, or MongoDB Atlas connectivity. The package also includes a vector retrieval ontology, a read-only vector collection inventory preset, and runtime configuration through AgentConfig, environment variables, and secret references.

The exposed MCP surface is intentionally compact: `vector_collection_management` handles collection-management operations, and `vector_search` handles search operations. A consolidated `vector-mcp-operations` workflow is available as the associated skill provider.

## How it works

The server uses action-routed tools to keep the MCP surface small and reduce context overhead. Its tool mode can be changed between intent, condensed, verbose, or both; the condensed mode presents the two action-routed tools listed above. Tool toggles can independently enable or disable collection management and search.

Operators can restrict visible tools or toolsets at startup with command-line arguments, environment variables, HTTP SSE headers, or HTTP SSE query parameters. Tool and tag filters can be resolved against known names, descriptions, and tags through a knowledge-graph layer, with a background cache refresh interval of 24 hours.

## Setup and configuration

Install the connector-focused package with:

```bash
uvx --from 'vector-mcp[mcp]' vector-mcp
```

Install only the provider extra needed by the deployment, such as `vector-mcp[postgres]`, `vector-mcp[qdrant]`, or `vector-mcp[mongodb]`. The `all` extra includes every optional provider along with the agent, Langfuse, and Logfire runtimes. The runtime requires `agent-connector-sdk>=0.2.0` and its full epistemic-graph engine contract.

For local IDE use, configure the process over stdio with `uvx`. For networked deployments, start it with `--transport streamable-http` and a port, or connect a client to an already deployed MCP URL. Supported deployment forms include stdio, Streamable HTTP, local container or `uv`, and a remote URL. Secret references such as database credentials and provider API keys can be passed through an alias-aware launcher; other launchers should resolve secrets through their own runtime boundary.

## Tools and capabilities

- `vector_collection_management`: exposes collection-management operations.
- `vector_search`: exposes search operations.
- Dynamic tool and toolset visibility filtering through CLI options, environment variables, SSE headers, and SSE query parameters.
- Provider selection for the native epistemic graph, PostgreSQL/pgvector, Qdrant, and MongoDB Atlas.
- Stdio and Streamable HTTP transports.

## Limitations and notes

Provider support is opt-in, and production images are intended to install only the providers they operate. The server does not expose a large fixed set of one-to-one MCP tools by default; its action-routed surface and selected tool mode determine what clients see. Network HTTP deployments require operator-managed host, proxy, and TLS policies; the generated configuration avoids an unauthenticated non-loopback listener. Detailed parameter schemas and validation constraints belong to the usage guide rather than the compact tool listing.

_Full upstream README: https://allmcps.com/mcp/vector-mcp/readme_