Skip to main content

Pluggable Tool-Call RBAC (MCP Governance)

⬅️ Back to Features Catalog

What It Does​

The Pluggable Tool-Call RBAC (MCP Governance) is a critical component of the LLM-Shield-Proxy. Intercepts autonomous JSON-RPC tool executions and enforces strict logical access controls against your existing Redis infrastructure. (OPA and HashiCorp Vault resolvers planned for v1.2)

How It Works​

This feature integrates directly into the zero-egress VPC architecture to ensure secure and ultra-low latency processing.

  1. Initialization: Configured during startup via policies.yaml or .env.
  2. Execution: Operates asynchronously within the data plane, guaranteeing high throughput.
  3. Completion: Mutates or validates the payload safely before egress to the upstream LLM provider.

View diagram on GitHub mobile 📱 -->

Performance Profile​

  • Execution Speed: Designed for microsecond-level latency impact.
  • Overhead: Highly concurrent execution without saturating the Python GIL.

Configuration Flags​

The engine operates automatically but can be tuned via deployment flags.

Environment Variable / ConfigDescriptionLinked Deployment Guide
ENABLE_PLUGGABLE_TOOL_CALL_RBAC_MCP_GOVERNANCEToggles this functionality.View in deployment.md

Critical Logic & Edge Cases​

  • Streaming Integrity: Seamlessly handles split token chunks in real-time.
  • Security Stance: Enforces a Zero-Trust, fail-closed default architecture.

FAQ​

Q: Does this break real-time streaming? A: No, the proxy is engineered to reconstruct and redact payloads on the fly without breaking SSE connections.

Q: Where can I see the audit logs for this feature? A: All decisions are exported via the Universal Decision Trace Exporter (OTel / OSCAL) for SOC 2 compliance.

Plainspeak​

This feature acts as a bouncer that strictly controls what an AI agent is allowed to do.

When an AI decides it wants to use a tool (like "delete a file" or "send an email"), it shouldn't be blindly trusted. This feature intercepts the AI's request before it happens, checks the AI's "ID badge" against a strict list of permissions, and blocks the action immediately if the AI isn't authorized to use that specific tool.

See the following test file for reference implementations and edge-case testing: tests/test_proxy.py.