Feature
Purpose-built security layer for Model Context Protocol. Validate every tool call, prevent prompt injection, block unauthorized access, and enforce granular policies across Claude, Cursor, and Windsurf agents.
MCP gives AI agents the ability to read files, query databases, make API calls, and execute code. Without a security layer, a single misconfigured tool can expose customer data, delete infrastructure, or exfiltrate secrets.
Malicious instructions embedded in tool outputs that hijack agent behaviour. Tork scans tool responses for injection patterns before they reach the model.
Agents calling tools they shouldn't have access to. Tork enforces tool allowlists and denylists per agent, per policy, per environment.
Sensitive data leaking through tool call parameters or responses. Tork scans all MCP traffic for PII and blocks or redacts before transmission.
Tool calls that reach internal services or make requests to attacker-controlled URLs. Tork validates URLs and blocks internal network access.
Every MCP tool call is intercepted before execution. The tool name, parameters, and context are captured for evaluation.
Parameters are validated against your policy rules. Input is scanned for PII, injection patterns, and malicious payloads. URLs are checked against allowlists.
Based on policy evaluation: allow the call, block it, redact sensitive parameters, or escalate to a human reviewer for approval.
Every decision generates an HMAC-signed compliance receipt with full context. Receipts are hash-chained for tamper detection.
Explicitly define which tools each agent can access. Everything else is blocked by default.
Type-check, range-check, and pattern-match every parameter against your schema.
50+ PII types detected in tool call inputs and outputs. Redact before the model sees it.
Per-tool, per-agent rate limits prevent abuse and runaway agents.
Block SSRF by restricting which URLs and domains tools can access.
Pattern matching for common prompt injection techniques in tool responses.
Emergency stop any agent instantly via API call or MCP tool.
HMAC hash-chained receipts for every tool call. Exportable for compliance.
Webhook notifications for blocked calls, PII detections, and policy violations.
Define security policies in YAML. Tork evaluates every MCP tool call against your rules in under 1ms.
policy.yaml
# MCP Security Policy
version: "1.0"
tools:
# Allow specific tools
allowed:
- read_file
- search_code
- run_tests
# Block dangerous tools
blocked:
- delete_file
- execute_shell
- send_email
# Require human approval
escalate:
- deploy_production
- modify_database
security:
# PII scanning on all tool calls
scan_pii: true
pii_action: redact # allow | redact | block
# URL restrictions
url_allowlist:
- "*.company.com"
- "api.github.com"
# Rate limits
rate_limit:
per_tool: 100/minute
per_agent: 500/minute
# Injection detection
scan_injections: true
injection_action: blockPython
from tork import TorkGovernance
tork = TorkGovernance(
api_key="tork_...",
policy="policy.yaml"
)
# Wrap any MCP tool call
result = tork.govern_tool_call(
tool="read_file",
params={"path": "/etc/passwd"},
)
if result.action == "block":
print(f"Blocked: {result.reason}")
else:
# Safe to execute
execute_tool(result.sanitized_params)JavaScript / TypeScript
import { TorkGovernance } from 'tork-governance';
const tork = new TorkGovernance({
apiKey: 'tork_...',
policy: 'policy.yaml',
});
// Wrap any MCP tool call
const result = await tork.governToolCall({
tool: 'read_file',
params: { path: '/etc/passwd' },
});
if (result.action === 'block') {
console.log(`Blocked: ${result.reason}`);
} else {
// Safe to execute
await executeTool(result.sanitizedParams);
}Tork's MCP security layer works as a transparent proxy. No changes to your MCP server code required — just point your client config at the Tork gateway.
Every tool call generates a blockchain-verified audit receipt. Export your complete governance history for compliance reporting and security audits.
Add security to every tool call in minutes. Free tier available.