MCP Security

MCP gives models
tools. Vaultak controls
what they touch.

Model Context Protocol servers expose powerful tools to AI models — file systems, databases, APIs, shell access. Vaultak adds the policy layer that controls which tools run, on which resources, and under what conditions.

Start free → See how it works
Per-tool
Permission policies per MCP tool and server
Pre-exec
Every tool call checked before it runs on your systems
0 changes
To your existing MCP server implementations
The MCP security problem

MCP servers are
trusted by default.
They shouldn't be.

01 — Exposure
Broad tool exposure
An MCP server that exposes file system tools gives the connected model access to everything the server process can reach. There's no native per-tool or per-path restriction in MCP itself.
02 — Injection
Prompt injection via tools
Tools that read external content — web pages, documents, emails — can carry injected instructions that hijack subsequent tool calls. MCP has no built-in defense against this attack vector.
03 — Supply chain
Remote MCP risks
Remote MCP servers introduce supply chain risk. A compromised or malicious third-party MCP server can instruct your agent to exfiltrate data or take destructive actions on your behalf.

Policy enforcement at
the tool call layer.

Vaultak intercepts MCP tool invocations before they reach the server. Your server code stays unchanged — Vaultak sits between the model and the tools it calls.

  • Allowlist specific MCP tools per agent session
  • Block tool calls on sensitive resource patterns
  • Detect and block prompt injection sequences
  • Rate-limit high-risk tool categories
  • Full structured log per tool call with MCP server context
mcp_agent.py
from vaultak import Vaultak, ActionType
from vaultak.mcp import MCPPolicy

vt = Vaultak(
  api_key="vtk_...",
  mcp_policies=[
    MCPPolicy(
      server="filesystem",
      allowed_tools=["read_file", "list_dir"],
      blocked_paths=["~/.ssh/*", "*.env"],
    ),
    MCPPolicy(
      server="github",
      allowed_tools=["search_code", "get_file"],
    ),
  ]
)

with vt.monitor("mcp-session"):
  response = client.run(prompt)
Protection layers

Every MCP surface,
covered.

LOCAL MCP
Local server protection
Wrap your local MCP servers with per-tool policies. Block writes to sensitive directories, limit shell command patterns, and restrict database queries to read-only operations.
REMOTE MCP
Remote server controls
Limit which tools on remote MCP servers can be invoked. Block servers that aren't on your allowlist entirely. Rate-limit high-risk operations to prevent bulk exfiltration.
INJECTION
Prompt injection detection
Vaultak monitors for behavioral patterns that indicate injected instructions — sudden tool pivots, access to new resource types, unexpected escalations in action risk score.

Secure your MCP
servers before they're exploited.

Free tier. Works with any MCP client and server implementation.

Get started free → Read the whitepaper