agenticvulhunter

AgenticVulHunter

AgenticVulHunter is a four-stage secure code review tool. It reviews your current Git changes and reports security issues that pass the validation threshold.

Step 1: Install

Install AgenticVulHunter using:

pipx install agenticvulhunter

Step 2: LLM Setup

You only need:

• Endpoint
• API key
• Model

You can configure them using avh_setup.toml or environment variables.

Option 1: avh_setup.toml

Create avh_setup.toml inside the repository you want to review:

[llm]

endpoint = "http://localhost:11434/v1"
api_key = ""
model = "qwen3-coder:30b"

Option 2: Environment Variables

export AVH_ENDPOINT="http://localhost:11434/v1"
export AVH_API_KEY="your-key"
export AVH_MODEL="qwen3-coder:30b"

Environment variables take priority over avh_setup.toml.

Step 3: Run

Run AgenticVulHunter inside your Git repository:

agenticvulhunter review

The terminal shows live progress for all four review stages. BM25 retrieval runs internally against the packaged static rule corpus, but is kept out of the user interface to avoid mixing internal ranking signals with validated findings.

The visible workflow follows the research pipeline:

1. Select potentially vulnerable changed lines.
2. Build repository knowledge around each line.
3. Search for matching CWE knowledge.
4. Validate the evidence and apply the requested threshold.

Every accepted finding also includes a repository-verified attack path. Source, propagation, control, and sink citations are checked against the referenced code before the finding is shown. If a citation or path link is inconsistent, the validator automatically sends a focused follow-up prompt; incomplete paths are kept in the run artifacts but are not promoted to findings.

The initial result shows the review and a small suggested code fix. In an interactive terminal, enter a finding number to reveal only its plain-language attack path, or press Enter to finish. Suggested fixes are never applied automatically.

To reopen the most recent result later, run:

agenticvulhunter findings

To browse all previous reviews and their final threshold-filtered findings in the terminal, run:

agenticvulhunter runs

Choose a run by number to see its final result. Enter another run number to switch runs, or use f# to open a finding's attack path. Source locations are clickable in terminals that support hyperlinks; d, a, and l open the reviewed diff, annotated diff, and run log.

The default validation threshold is 0.6.

To use another threshold:

agenticvulhunter review 0.7