Give AI agents access to enterprise knowledge

AI agents need controlled access to company knowledge. The VectorAmp MCP service exposes dataset, search, document, Intelligence, and ingestion tools through Streamable HTTP so agents can query governed VectorAmp datasets instead of relying on pasted context or uncontrolled file access.

This guide shows how to connect common MCP clients, including Cursor, Claude Desktop, OpenAI Codex, Claude Code, and clients that support Streamable HTTP directly.

What you will build

An MCP client configuration that lets an agent:

list VectorAmp datasets
search a dataset with metadata filters
ask cited Intelligence questions
inspect and download retained source documents
run ingestion/source/job tools when its credential allows it

Prerequisites

Access to a deployed VectorAmp MCP endpoint.
A VectorAmp access token or API key authorized for the dataset.
Node.js 20.18.1+ or a current LTS release if your client uses mcp-remote.
An MCP-compatible client.

Endpoint and transport

Production endpoint:

https://mcp.vectoramp.com/mcp

Use Streamable HTTP. Do not configure legacy SSE.

Many desktop coding tools still launch MCP servers over local stdio. For those clients, use mcp-remote as a bridge from local stdio to VectorAmp's remote Streamable HTTP endpoint.

npx -y mcp-remote@latest --help

Authentication options

VectorAmp MCP supports two auth modes.

Method	Best for	How it is sent
OAuth / browser login	human-operated desktop clients	`Authorization: Bearer <access token>`
API key	headless agents, CI, simple local setup	`X-API-Key: vsk_...`

For API-key examples, create a VectorAmp API key in Settings → API keys and keep it in an environment variable:

export VECTORAMP_API_KEY="vsk_..."

tip

For desktop clients launched from Finder, Spotlight, Dock, or Start Menu, shell environment variables may not be inherited. If the client cannot see VECTORAMP_API_KEY, either use the client's env block or a small wrapper script as shown below.

Cursor

Cursor supports stdio-launched MCP servers, so use mcp-remote.

{
  "mcpServers": {
    "vectoramp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote@latest",
        "https://mcp.vectoramp.com/mcp",
        "--transport",
        "http-only"
      ]
    }
  }
}

The first connection should open the VectorAmp browser login flow.

Cursor with an API key

{
  "mcpServers": {
    "vectoramp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote@latest",
        "https://mcp.vectoramp.com/mcp",
        "--transport",
        "http-only",
        "--header",
        "X-API-Key:${VECTORAMP_API_KEY}"
      ],
      "env": {
        "VECTORAMP_API_KEY": "vsk_..."
      }
    }
  }
}

The no-space X-API-Key:${...} form avoids argument escaping issues in some MCP clients.

Claude Desktop

Claude Desktop also uses local stdio MCP servers, so the setup is almost identical to Cursor.

Open Claude Desktop's MCP config file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\\Claude\\claude_desktop_config.json

{
  "mcpServers": {
    "vectoramp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote@latest",
        "https://mcp.vectoramp.com/mcp",
        "--transport",
        "http-only"
      ]
    }
  }
}

Fully quit and restart Claude Desktop after editing the file.

Claude Desktop with an API key

{
  "mcpServers": {
    "vectoramp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote@latest",
        "https://mcp.vectoramp.com/mcp",
        "--transport",
        "http-only",
        "--header",
        "X-API-Key:${VECTORAMP_API_KEY}"
      ],
      "env": {
        "VECTORAMP_API_KEY": "vsk_..."
      }
    }
  }
}

If Claude Desktop cannot find npx, use an absolute path to a modern npx or wrap the command in a script that sets PATH before launching npx.

OpenAI Codex CLI

OpenAI Codex can add MCP servers from the command line. Use mcp-remote to bridge Codex to VectorAmp's Streamable HTTP endpoint.

codex mcp add vectoramp -- \
  npx -y mcp-remote@latest \
  https://mcp.vectoramp.com/mcp \
  --transport http-only

Start Codex after adding the server. The first tool call should trigger the browser OAuth flow if the client has no cached token.

Codex with an API key

For API keys, prefer a wrapper script so the key stays in your shell or secret manager instead of being hard-coded into the Codex MCP command.

Create ~/bin/vectoramp-mcp:

#!/usr/bin/env bash
set -euo pipefail

exec npx -y mcp-remote@latest \
  https://mcp.vectoramp.com/mcp \
  --transport http-only \
  --header "X-API-Key:${VECTORAMP_API_KEY}"

Make it executable and register it with Codex:

chmod +x ~/bin/vectoramp-mcp
export VECTORAMP_API_KEY="vsk_..."

codex mcp add vectoramp -- ~/bin/vectoramp-mcp

When you launch Codex, make sure VECTORAMP_API_KEY is present in the environment.

Claude Code CLI

Claude Code can also run local stdio MCP commands. Use the same bridge pattern.

claude mcp add vectoramp -- \
  npx -y mcp-remote@latest \
  https://mcp.vectoramp.com/mcp \
  --transport http-only

Claude Code with an API key

Use the same wrapper script from the Codex section, then register it:

export VECTORAMP_API_KEY="vsk_..."
claude mcp add vectoramp -- ~/bin/vectoramp-mcp

If your Claude Code installation uses a JSON config instead of the claude mcp add helper, use the same mcpServers JSON shape shown for Claude Desktop.

Direct Streamable HTTP clients

If your MCP client supports remote Streamable HTTP directly, do not use mcp-remote. Configure the URL and transport directly.

{
  "mcpServers": {
    "vectoramp": {
      "url": "https://mcp.vectoramp.com/mcp",
      "transport": "streamable-http"
    }
  }
}

With an API key:

{
  "mcpServers": {
    "vectoramp": {
      "url": "https://mcp.vectoramp.com/mcp",
      "transport": "streamable-http",
      "headers": {
        "X-API-Key": "vsk_..."
      }
    }
  }
}

Other stdio MCP clients

For clients such as Windsurf, Cline, Continue, Zed, or custom local agent runners, use the same stdio bridge shape if they accept command, args, and optional env:

{
  "mcpServers": {
    "vectoramp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote@latest",
        "https://mcp.vectoramp.com/mcp",
        "--transport",
        "http-only",
        "--header",
        "X-API-Key:${VECTORAMP_API_KEY}"
      ],
      "env": {
        "VECTORAMP_API_KEY": "vsk_..."
      }
    }
  }
}

If the client supports OAuth through mcp-remote, remove the --header and env lines and let the browser login flow run.

Test the connection

After connecting the client, ask the agent to enumerate tools or run a harmless read-only query.

List the VectorAmp MCP tools you can use. Then list my datasets and stop.

Useful tools include:

list_datasets
get_dataset
search_dataset
ask_vectoramp
list_dataset_documents
download_dataset_document
list_ingestion_jobs

Example agent instruction:

Use VectorAmp to search the product-docs dataset for "S3 ingestion timeout".
Return a short answer with cited source titles and document IDs.

Search with filters

The search_dataset tool supports text or vector search, top_k, metadata filters, advanced filters, hybrid search, and reranking controls.

Example tool arguments:

{
  "dataset_id": "ds_123",
  "text": "S3 ingestion timeout",
  "top_k": 8,
  "filters": {
    "product_area": "ingestion",
    "status": "current"
  },
  "hybrid": true,
  "sparse_query": "S3 timeout",
  "rerank_enabled": true
}

Troubleshooting

Symptom	What to check
Client tries SSE	Force `--transport http-only` for `mcp-remote`, or choose `streamable-http` in direct clients.
Browser login never opens	Confirm Node.js is current and the endpoint includes `/mcp`.
`npx` not found in Claude Desktop	Use an absolute `npx` path or a wrapper script that sets `PATH`.
API key appears ignored	Prefer `X-API-Key:${VECTORAMP_API_KEY}` with no space, and verify the client passes `env`.
Agent sees no datasets	Confirm the token/API key has access to the expected organization and dataset.

Production guardrails

Use least-privilege API keys or OAuth scopes.
Prefer dataset-specific access.
Keep sensitive datasets out of broad agent workspaces.
Audit tool usage where possible.
Require citations for decision-making answers.
Do not commit API keys into shared MCP config files.

What you will build​

Prerequisites​

Endpoint and transport​

Authentication options​

Cursor​

Cursor with OAuth/browser login​

Cursor with an API key​

Claude Desktop​

Claude Desktop with OAuth/browser login​

Claude Desktop with an API key​

OpenAI Codex CLI​

Codex with OAuth/browser login​

Codex with an API key​

Claude Code CLI​

Claude Code with OAuth/browser login​

Claude Code with an API key​

Direct Streamable HTTP clients​

Other stdio MCP clients​

Test the connection​

Search with filters​

Troubleshooting​

Production guardrails​

Related docs​