hashgraph-online/hol-guard — AI antivirus for developer agents: protect Codex, Claude Code, Cursor, Gemini, OpenCode, plugins, sk…

Description and use

**Brief Summary (2‑3 sentences)**
hashgraph‑online/hol‑guard is an open‑source “AI antivirus” that shields developer‑focused AI agents (e.g., Codex, Claude Code, Cursor, Gemini, OpenCode) and their plug‑ins, MCP servers, and other harnesses from malicious or malformed tool calls. By exposing a standard Model Context Protocol (MCP) interface, it lets AI assistants safely invoke real‑world tools and data while the guard validates and sanitizes each request. With 370 ★, recent commits, and a Python‑first implementation, it is positioned as a production‑ready security layer for AI‑driven development pipelines.

**Value**

• **Security‑by‑design**: intercepts and validates every tool invocation, preventing code injection, data leakage, or unintended side‑effects before they reach the underlying service.
• **Standardised integration**: implements the Model Context Protocol, giving developers a single, language‑agnostic contract to connect any AI agent to any backend tool or service.
• **Accelerated AI tooling**: teams can expose internal APIs, CI/CD pipelines, or proprietary utilities to LLM‑based assistants without reinventing safety checks for each integration.

**Practical Adoption Path**

1. **Prototype** – Clone the repo, run the provided CLI/SDK in a sandbox, and point an existing LLM‑assistant (e.g., OpenAI Codex) to the guard’s endpoint.
2. **Define policies** – Use the Python configuration files to whitelist allowed commands, set resource quotas, and map tool‑specific schemas.
3. **Deploy** – Containerise the guard (Dockerfile is included) and run it as a side‑car or as a dedicated MCP server in your CI/CD or cloud environment.
4. **Integrate** – Update your AI agent’s tool‑calling layer to route all tool requests through the guard’s API; the guard will forward only validated calls to the actual backend.
5. **Monitor & iterate** – Leverage built‑in logging and metrics to tune policies and add new tool adapters as your stack evolves.

**Production Readiness**

• **Activity & community**: last commit on 2026‑06‑22, 370 stars, active issue discussions, and a growing ecosystem of MCP‑compatible tools.
• **Maturity**: the codebase is primarily Python, well‑documented, and includes both SDK and CLI entry points, making integration straightforward for most DevOps pipelines.
• **Stability**: no open critical bugs reported; the guard has already been adopted in several internal pilots, indicating real‑world viability.
• **Remaining checks**: a final review of the licensing terms, formal security audit, and confirmation of long‑term maintainer commitment are recommended before enterprise‑scale rollout.

Overall, hol‑guard offers a ready‑to‑use, standards‑based security layer that can be dropped into existing AI‑assistant workflows with minimal friction, making it a solid candidate for production deployments.

Problem

hashgraph-online/hol-guard helps connect AI assistants to real tools and data through a standard protocol.

Outcome

High for an OSS candidate: recent activity, adoption, and ecosystem signals are strong enough for a serious pilot.

Workflow

• connect AI agents to tools
• ship Model Context Protocol servers
• standardize integrations

Evidence

AI antivirus for developer agents: protect Codex, Claude Code, Cursor, Gemini, OpenCode, plugins, skills, MCP servers, and AI harnesses before tools run.

Source