Nacos 3.2 Skill Registry Launches to Enable Secure, Controllable Enterprise AI

Over the past year the internal team debated whether the future belongs to low‑code platforms (e.g., Dify, low barrier but limited flexibility) or high‑code frameworks built on the ReAct structure (LLM + Prompt + Tool, highly flexible but prone to hallucinations and costly prompt debugging). Anthropic’s Skill emerged as a middle ground, retaining flexibility while encapsulating capabilities and imposing boundary constraints, leading to rapid adoption and a Skill market with tens of thousands of entries.

A security analysis by Snyk on a sample of 3,984 Skills from ClawHub revealed that 36.82% contain vulnerabilities, 13.4% are classified as CRITICAL, and issues such as credential leaks and prompt injection are common, posing major challenges for enterprise deployment of Agents.

Four intertwined challenges were identified from discussions with internal and external teams: Security (malicious code, known bugs, data leakage), Permission (unclear visibility and access rights), Stability (version chaos, uncontrolled upgrades, difficult rollbacks), and Governance (missing audit logs and compliance evidence). Solving any single issue is insufficient; a platform‑level governance system is required.

Nacos 3.2 introduces a private Skill Registry designed for production environments. It creates a “verify‑then‑trust” layer between Agents and Skills, consolidating audit, management, distribution, and traceability into a complete enterprise‑grade AI resource control loop.

Quick Start – Install Nacos with a single script:

curl -fsSL https://nacos.io/nacos-installer.sh | bash

or on Windows PowerShell:

iwr -UseBasicParsing https://nacos.io/nacos-installer.ps1 | iex

0. Skill Security Audit Pipeline – Nacos 3.2 ships an out‑of‑the‑box security audit plugin that scans for more than ten common risks and provides a standard interface for custom plugins. The pipeline enforces a “not passed, not released” rule, turning security requirements from documentation into system‑enforced constraints, and performs multi‑dimensional scanning, risk grading, and automated decision making before a Skill is stored.

1. Multi‑Version Management & Gray Release – Supports a full lifecycle (DRAFT, REVIEWING, GRAY, FORMAL, OFFLINE) with immutable versions. Tag‑based labels (dev/latest/stable) enable precise rollout control and gradual traffic shifting. If an issue is discovered during gray testing, switching the label rolls back the version instantly, preserving business continuity.

2. Permission Model & Multi‑Layer Isolation – Implements three‑layer boundaries: RBAC roles (publisher, auditor, read‑only) define clear responsibilities; namespace isolation separates teams, environments, and tenants; Skill‑level visibility configures each Skill as public, private, or limited‑access, ensuring that “who can see, use, modify, or publish” is explicitly governed.

3. Full‑Link Audit & Traceability – Records every operation: upload logs (who uploaded which version and when), audit logs (who approved, decision, timestamp), release logs (who released to which Agents and under what policy), and call trace (which Agent invoked which version). This audit closure enables operations, compliance, and governance to form a verifiable system.

4. Skill Access & Distribution – Agents can discover Skills automatically via a curl request, e.g. curl -s https://download.nacos.io/SKILL.md, which returns installation guidance and a skill list. The Nacos CLI provides one‑click commands such as nacos-cli skill-list and nacos-cli skill-get mysql-query -o ~/.skills. Batch acquisition is supported through a skills.sh script that respects the SKILLS_API_URL environment variable, fitting CI/CD pipelines.

Open Architecture – The registry is plugin‑able, allowing enterprises to extend functionality. It integrates with HiClaw (Agent runtime that consumes Skills, Prompts, AgentSpecs) and HiMarket (a private Skill and Worker marketplace built on Nacos), forming a complete enterprise AI management ecosystem.

Future Roadmap – Planned enhancements include AI resource lifecycle management with activity‑based loading, a data‑intelligence layer that ingests OT‑protocol tracing for Copilot‑style analysis, enriched full‑link audit, semantic search powered by vector databases and large models, support for additional Agent communication protocols (ACP, Matrix), and a Coding Agent plugin that evolves from Markdown commands to native MCP endpoints.

In summary, Nacos 3.2 marks the first step toward a unified AI resource control plane, aiming to become the foundational infrastructure for multi‑Agent collaboration in the AI era.

References: Snyk report on Skill vulnerabilities; GitHub Issue #14770 (https://github.com/alibaba/nacos/issues/14770).