Vaultak MCP Server

AI agent runtime security — directly inside Claude, Cursor, and any MCP-compatible assistant.

What it does

The Vaultak MCP server gives any AI assistant three security superpowers:

Install

pip install vaultak-mcp

Or run directly:

uvx vaultak-mcp

Connect to Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "vaultak": {
      "command": "uvx",
      "args": ["vaultak-mcp"],
      "env": {
        "VAULTAK_API_KEY": "your_api_key_here"
      }
    }
  }
}

Get your API key at app.vaultak.com.

Connect to Cursor

In Cursor Settings → MCP → Add Server:

{
  "name": "vaultak",
  "command": "uvx vaultak-mcp"
}

Example usage

Once connected, ask your AI assistant:

"Score the risk level of my agent that has access to the production database and can send emails"

"Should my agent be allowed to delete records from the users table?"

"Give me policy templates for protecting PII in my AI agent"

Tools reference

vaultak_risk_score

agent_description  (required)  What the agent does
capabilities       (optional)  List of tools/capabilities
api_key            (optional)  Vaultak API key for live scoring

Returns composite score (0–100), risk tier, per-dimension breakdown, and recommendations.

vaultak_policy_check

action    (required)  The action being attempted (delete, write, execute...)
resource  (required)  The resource being acted on (production_db, /etc/secrets...)
policies  (optional)  Policy list to evaluate against
api_key   (optional)  Fetch your live policies automatically

Returns ALLOW or BLOCK decision with matched policy and recommendation.

vaultak_get_policy_templates

scenario  (required)  One of: database_protection | file_system_limits |
                               api_rate_limiting | pii_protection |
                               production_safeguards | all

Returns ready-to-use policy JSON for your Vaultak dashboard.

Links

• Dashboard: app.vaultak.com
• Docs: docs.vaultak.com
• SDK: pip install vaultak
• Site: vaultak.com