agentgg-agents

The official agent library for agentgg — AI-powered SAST agents for code security review.

Every template is one kind of thing: an agent. A .md file with YAML frontmatter (a precondition and a where) and a markdown body that is the prompt. There are no execution modes — every agent is a tool-enabled investigation (Read/Glob/Grep) that runs over the files its where selects.

agentgg downloads this catalog automatically on first scan and refreshes it with agentgg agents update. No manual setup needed.

Documentation · agentgg CLI · agentgg.dev

Directory structure

agentgg-agents/
├── agents/           # The agent library, organized by category
│   ├── injection/    # SQL, NoSQL, command, XSS, path traversal, mass assignment
│   ├── auth/         # Authentication, authorization, JWT, OAuth, session, IDOR
│   ├── misconfiguration/ # CORS, caching, cookies, feature-flag security
│   ├── logic/        # Race conditions, async bugs, event handler mismatches
│   ├── infrastructure/ # Docker, Kubernetes, Terraform, GitHub Actions
│   ├── cloud/        # AWS Lambda, GCP, Azure, IAM
│   ├── cryptography/ # Insecure algorithms, unsafe deserialization
│   ├── mobile/       # Android, iOS
│   ├── smartcontract/ # Solidity access control, reentrancy
│   ├── ai/           # Agent loops, tool definitions, MCP handlers
│   └── deep/         # Broad-net variants — opt-in only, never run by default
└── semgrep-rules/    # Shared semgrep rules agents reference by name

Every category except deep/ runs when no -t flag is given. deep/ agents cast a much wider net per run, so they are opt-in via -t.

Usage

agentgg scan ./src                              # every category except deep/
agentgg scan ./src -t agents/injection/         # one category
agentgg scan ./src -t agents/injection/ -t agents/auth/   # multiple
agentgg scan ./src -t agents/deep/              # the broad-net variants
agentgg scan ./src -t sql-injection             # a single agent by slug
agentgg agents update                           # refresh the catalog

See Choose agents for the full selection rules.

Write an agent

An agent declares a precondition (should this run on this repo?), a where (which files?), and the instructions (the prompt body):

---
slug: sql-injection
name: SQL Injection
description: SQL built from untrusted input instead of parameterized queries.
version: 0.1.0
author: your-github-handle
noiseTier: normal
precondition:
  regex:
    patterns:
      - regex: "\\.(query|execute)\\s*\\("
        in: ["**/*.{ts,js,py,go,php}"]
where:
  extensions: [ts, js, py, go, php]
  excludePatterns: ["**/*.{test,spec}.*"]
  preFilter:
    - { regex: "\\.(query|execute)\\s*\\(", label: "raw SQL call" }
references:
  - CWE-89
  - OWASP-A03:2021
---

You are reviewing source code for SQL injection...

Every frontmatter field is documented at:

• Agent anatomy — the file format and every frontmatter field
• Targeting — precondition and where in full
• Create from reports — turn a past incident into an agent
• Manage agents — list, lint, and update the catalog

Contributing

See CONTRIBUTING.md for the full guide. In short:

1. Fork this repo
2. Add a new .md agent in the appropriate category folder (slug must match the filename)
3. Test it locally: agentgg scan ./test-fixture -t ./your-agent.md
4. Lint the whole tree: agentgg agents lint .
5. Open a pull request

New agents should have:

• A focused, single-responsibility prompt with clear true-positive and false-positive criteria
• A where scoped to the relevant files (extensions plus a preFilter for the suspicious construct)
• A precondition so the agent skips repos it can't apply to (cheap regex, and/or a prompt gate)
• A CWE / CVE / OWASP reference when one fits (optional)

To report a vulnerability privately, see SECURITY.md.

License

The agent library is licensed under the MIT License. See LICENSE for the full text.