AgentShield
https://github.com/affaan-m/agentshield- Category
- Cybersecurity
- Rank
- No. 624Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- affaan-m
- GitHub
- 1.2k stars
- Latest release
- v1.4.0
- Date
About
AI agent security scanner that detects vulnerabilities in agent configurations, MCP servers, and tool permissions. Ships as a CLI, GitHub Action, and GitHub App.
What it does
AgentShield reads agent configuration and instruction files, then applies static rules to flag exposed credentials, broad permissions, unsafe hooks, risky server definitions, prompt injection patterns, and package-manager weaknesses. It separates active settings from templates and plugin caches, offers selected automatic fixes, and produces terminal, JSON, Markdown, HTML, SARIF, or portable evidence output.
Why it's ranked here
The strongest case is breadth joined to workflow fit. Findings carry severity, source context, remediation guidance, and optional stable fingerprints. Baseline comparison, policy enforcement, supply-chain checks, evidence bundles, and CI annotations turn detection into an enforceable process. The repository also declares extensive automated test batches. Its focus remains agent configuration security, especially Claude Code, rather than general application security.
What's good
Source context is unusually useful. AgentShield distinguishes active configuration, optional local settings, examples, documentation, plugin material, and hook code. It reduces the weight of template findings so shipped examples do not look fully enabled. Safe remediation can replace hardcoded secrets with environment references, while initialization creates scoped permissions, explicit denials, safety hooks, and an empty server configuration without overwriting existing files.
Tradeoffs
The default scope and generated baseline center on Claude Code, even though reports can record markers from other development harnesses. Deeper modes add operational cost: model-assisted analysis requires an Anthropic API key, online supply-chain inspection needs network access, and sandbox or injection modes execute more than static checks. Rule findings still require human confirmation, especially for templates and broad risk classifications.
How to use it well
Use AgentShield before accepting shared agent configuration, when adding hooks or external tool servers, and in CI to block new critical or high-risk findings against a saved baseline. Security teams can add organization policies and retain evidence bundles for audits. Treat automatic fixes as a starting point, then review permissions manually. Pair it with application security testing and runtime monitoring, which its configuration scans do not replace.
Technical notes+
package.json defines an ESM TypeScript package for Node.js 18 or newer, exposes the agentshield binary and importable package entry, and builds with tsup. src/index.ts wires Commander commands to scanning, severity filtering, report renderers, baselines, policies, evidence packs, supply-chain analysis, optional active analysis, fixes, initialization, and watching. src/action.ts implements GitHub Actions integration through native workflow commands, annotations, outputs, summaries, and failure gates without @actions/core. src/types.ts defines findings, runtime-confidence classes, reports, harness adapters, and deep-analysis results. src/rules/mcp.ts assigns source confidence and downgrades template severity. src/fingerprint.ts hashes evidence into stable finding fingerprints. src/init/index.ts creates baseline configuration while skipping existing files.
Observed
- License
- MIT
- Primary language
- TypeScript
- Runtime
- Node.js 18 or newer
- Packaging
- ESM npm package named ecc-agentshield, runnable through npx or global installation
- Interfaces
- Command-line binary, importable package, GitHub Action, and GitHub App integration
- Report formats
- Terminal, JSON, Markdown, HTML, and SARIF, plus portable evidence bundles
- Testing structure
- Vitest batches cover rules, scanning, reporting, integration, actions, policies, supply chain, fixes, sandboxing, runtime, baselines, and evidence packs
Read from README.md, package.json, src/index.ts, src/types.ts, src/action.ts, src/fingerprint.ts, src/action-policy.ts, src/source-context.ts, src/action-baseline.ts, src/action-hardening.ts, src/action-promotion.ts, src/action-supply-chain.ts, src/rules/mcp.ts, src/init/index.ts, src/opus/index.ts.
What it can do
Scan AI agent configurations for security vulnerabilities
AI agent configuration files → Vulnerability report with detected security issues
Analyze MCP server security settings
MCP server configuration → Security assessment report
Audit tool permissions for potential risks
Tool permission configurations → Permission risk analysis report
Run security scans via command line interface
CLI commands and target files → Security scan results in terminal
Automate security checks in GitHub workflows
GitHub repository and workflow triggers → Automated security scan results in GitHub Actions
Monitor repository security continuously
GitHub repository with AgentShield app installed → Ongoing security monitoring and alerts
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.