The problem

‍

When an agent sends a prompt or invokes tools/call, secrets and PII can leave your runtime with no policy decision and no audit record. Prompt engineering does not scale. CASB was not built for agent-to-LLM HTTPS.

‍

‍

How Blekline handles it

‍

• MCP proxy — every downstream tools/call passes through mask, enforce, or block before execution.
• Trust Vault — tokenize secrets before LLM context; hydrate for approved tool calls inside your VPC only.
• Ingress API — OpenAI-compatible base URL swap for SDK and sidecar routes.
• Stateful masking — placeholders persist across multi-turn sessions.
• Workspace policy — YAML and contracts via @blekline/contracts.

‍

‍

What ships today

‍

SurfacePackage / routeStatusMCP proxy@blekline/mcp-proxyShippedMCP server@blekline/mcp-serverShippedIngress API/api/ingress/v1/*ShippedNHIM sidecarghcr.io/blekline/sidecarShipped

‍

‍

Flow

‍

Agent runtime → Blekline (mask / enforce / block) → model API or downstream MCP.

‍

‍

Use case

‍

An integrator ships a customer-care agent. Blekline ensures the agent cannot exfiltrate a full customer database to an external endpoint via tool calls — whether by user mistake or prompt injection.

‍

‍

Honest scope

‍

Blekline governs agent interactions at the MCP, SDK, and API boundary — not every SaaS app or network packet. Pair ingress control with execution and audit for a full Layer 4 program.