CCS Protocol Integration - TypeScript SDK for Agent Governance

[Correctover]

Hi inkeep team!

Love the "No-Code Visual Builder or TypeScript SDK" approach. Want to introduce CCS (Correctover Conformance Standard) — a protocol that can enhance your TypeScript SDK with framework-level security guarantees.

The Problem: CWE-636

We discovered that all major Agent frameworks (CrewAI, AutoGen, LangGraph) use observer-pattern hooks that fail-open when governance throws exceptions:

Hook: governance_crash → exception caught → hook_blocked=False → tool EXECUTES ❌

CVSS 9.1, affecting the entire ecosystem.

CCS Solution

CCS uses interceptor decorators instead of hooks:

CCS: governance_crash → exception caught → tool NEVER CALLED ✅

TypeScript SDK

We just released the TypeScript version (npm package):

npm install correctover-ccs

import { govern } from "correctover-ccs";

const myTool = govern(
  async (args: any) => {
    return "result";
  },
  { policy: "default" }
);

Governance failures structurally block tool execution. Not "logged and allowed" — blocked.

Integration with inkeep

inkeep's TypeScript SDK can leverage CCS for:

1. Framework-level security: Interceptor pattern ensures fail-closed
2. Zero-friction integration: 2 lines of code
3. Cross-framework compatibility: Works with any framework

CCS Verified Certification

We're offering CCS Verified certification for projects using interceptor-based governance.

inkeep would be an ideal candidate, especially given your focus on enterprise-grade agent platforms.

Resources

• npm: https://www.npmjs.com/package/correctover-ccs
• SDK: https://github.com/Correctover/ccs-sdk
• Protocol: https://github.com/Correctover/standards
• Paper: DOI 10.5281/zenodo.21234580
• CVE Analysis: https://gist.github.com/Correctover/9cfb97bcf374f79b793fd0bacd4e9d62

Interested in exploring integration or certification?

Thanks!

[pullfrog]

  ^{｜ Build this ➔ ｜ Make a plan ➔}

[hegu-1]

The fail-open vs fail-closed distinction is the part I would keep front and center here.

For production agent systems, a governance check should be treated like an execution precondition, not an observer hook. If the policy layer crashes, times out, or cannot classify the action, the tool call should not happen by default. Otherwise the governance layer becomes documentation rather than control.

A useful implementation contract might include:

• policy result is required before tool execution
• unknown action class defaults to ask/deny
• denied/skipped calls produce an audit event
• tool output can be firewalled before returning to the model
• tests intentionally crash the governance layer and assert no side effect occurred

I am mapping this kind of control-plane boundary in an enterprise AI OS architecture here: https://github.com/hegu-1/enterprise-ai-os-architecture