MCP ingress · bounded runtime verified

Trust policy before
tool execution.

MCP Hub provides a server-backed trust-policy control plane, append-only security evidence, bounded guarded execution, and a stateless MCP ingress live-verified against the isolated proof tool.

Current claim boundary: A real external MCP client reached the deployed Base44 ingress and reconciled ALLOW evidence to independent `EnforcementProbe` state. External-provider enforcement, the complete RLS matrix, frontend certification, and production certification remain blocked.
Open the Control Plane View Enforcement Status
Versioned
Trust policy
Append-only
Security evidence
Fail-closed
Security state
Bounded
Execution proof

Preview the verdict model.

Simulated examples of ALLOW, FLAG, and BLOCK outcomes. No live request is evaluated on this public page.

mcp-hub · trust evaluator · SIMULATEDsimulating book-annotator-prod
Request fingerprint
Agent identity verified
Scope within policy
Behavior pattern normal
Anomalous call frequency
Prompt injection attempt

Legitimate agent

Identity verified · Scope compliant · Normal behavior

Unknown agent

No registration · Attempted tool enumeration

Known agent — anomaly

Identity ok · Call rate 40× baseline spike

Prompt injection attempt

Jailbreak pattern detected in tool argument

How the verified enforcement boundary works

This flow is live-verified through the deployed Base44 endpoint for the bounded `probe.increment` proof tool. External-provider execution and production certification remain separate gates.

  1. Request enters the MCP boundary

    The implemented ingress authenticates a bearer credential, derives client identity server-side, validates MCP routing metadata, and exposes only bounded registered tools.

  2. Server policy decides

    The latest versioned policy evaluates the requested tool and arguments and returns ALLOW, FLAG, or BLOCK with reason codes.

  3. Execution is conditional

    Only ALLOW reaches the bounded mutation callback. FLAG requires review and BLOCK stops execution. Every decision can produce immutable evidence.

What we evaluate on every call

Six layers of evaluation fire in parallel on every inbound agent request — under 5ms, no rule writing.

  • Agent Identity Verification

    Each agent registers a cryptographic identity. Every request is verified against that identity before any tool is exposed.

  • Behavioral Anomaly Detection

    Hub tracks call patterns, tool enumeration, and frequency spikes. Deviations from baseline trigger automatic review.

  • Prompt Injection Scanning

    Incoming tool arguments are scanned for injection patterns, jailbreak payloads, and policy-violating instructions.

  • Least-Privilege Scoping

    Each agent is granted only the tools it declared it needs. Scope creep is blocked before execution.

  • Real-Time Threat Logging

    All evaluation decisions — pass or block — are written to an immutable audit log with full request context.

  • Risk Scoring Dashboard

    Per-agent risk scores surface trending anomalies, repeat offenders, and tool abuse patterns in one view.

Know which agents to trust.
Before they act.

Stop hoping your agents behave. Connect them through MCP Hub and get a real-time verdict on every request — legitimate, suspicious, or malicious.

Get Started Free Read the Docs
  • No credit card required
  • Set up in minutes
  • MCP-compliant