v0.4.0GitHub

Core concepts

AI Enablement Stack

Five layers from infrastructure to autonomous agents — where Blekline and L1 sandboxes fit.

View as Markdown

Enterprise AI adoption stacks five layers. Blekline operates at Layer 4 — Non-Human Identity & Runtime Enforcement (MCP ingress and sidecar). Layer 1 sandbox providers (Daytona, Modal, Vercel Sandbox, Cloudflare, E2B) run isolated agent code.

The five layers

LayerNameRoleExamples
L5Agent consumerAutonomous and assistive agents that plan, call tools, write codeSee L5 table below

L5 agent clients (OSS config paths)

ClientCapabilityOSS pathDocs
CLI / SDKverifiedcli/Quick start
Cursorverified.cursor/mcp.json.exampleCursor
Claude Codeverified.claude/settings.json.exampleClaude Code
Claude Desktopverifiedconfig/claude_desktop_config.json.exampleClaude Desktop
Codexverified.codex/config.toml.exampleCodex
GitHub Copilotverified.vscode/mcp.json.exampleCopilot
Continueverified.vscode/continue.config.json.exampleContinue
OpenHandspattern— (app docs)OpenHands
Sourcegraph Codypattern— (app docs)Cody
Claude connectorcloud-onlyremote MCP on app.blekline.comClaude connector
L4Runtime enforcement & NHIMMask, classify risk, enforce MCP tool policy, auditBlekline — security plane for machine identities
L3EngineeringTraining, eval, testing, quality for agent systemsLangSmith, Guardrails, Braintrust
L2IntelligenceFrameworks, retrieval, coding modelsLangChain, LlamaIndex, Azure OpenAI, OpenRouter
L1InfrastructureSandboxes, model APIs, cloud computeSandbox providers — Daytona, Modal, E2B, Cloudflare, Vercel

Request flow (Blekline + L1 sandbox)

flowchart TB
  subgraph L5["L5 — Agent consumer"]
    A[Cursor / Claude / Codex]
  end
  subgraph L4["L4 — Blekline governance"]
    MS[mcp-server]
    MP[mcp-proxy]
    CP[mask · enforce · audit]
  end
  subgraph L1a["L1 — Sandbox MCP"]
    S[L1 runtime MCP]
    M[Model APIs]
  end
  A --> MS
  A --> MP
  MP --> MS
  MS --> CP
  MP --> CP
  CP --> M
  MP --> S

Why Layer 4 matters

Teams ship L5 agents before L4 governance is in place. One tool_call can leak an API key, run an unapproved shell command, or send PII to a model — while the agent UI still looks healthy.

Blekline sits at the MCP boundary: every prompt and tool invocation can be masked, evaluated, and audited before traffic reaches L1 runtimes or model APIs.

See Why ingress governance for the problem framing and Architecture for component wiring.

Open core vs cloud

CapabilityOSS (blekline-oss)Cloud (app.blekline.com)
MCP server / proxyYesYes
Local secret + tool enforceYes (@blekline/contracts)Yes
Azure authoritative PII maskYes
Fleet policy (SSE)Yes
Investigations / billingYes

Next steps: Quick start · Agent clients · Model providers · Sandbox providers · MCP Cursor setup · Open workspace