Documentation

CLI-first assurance for AI agents.

GuardPrompt is designed to start as a developer-first scanner for AI agent projects. The CLI validates behavior contracts, detects risky capabilities, runs adversarial tests, and produces deployment evidence.

Planned quick start

npm install -g guardprompt

guardprompt init
guardprompt scan ./agent
guardprompt report --format html

CLI package is planned. Current site documents the product direction, security model, and initial interface.

Initialize a contract

guardprompt init

Creates a starter guardprompt.yaml behavior contract for an agent project.

Scan an agent

guardprompt scan ./agent

Analyzes agent files, tools, prompts, memory configuration, and behavior contract.

Block critical findings

guardprompt scan . --fail-on critical

Fails CI/CD when critical policy violations are detected.

Generate evidence

guardprompt report --format json

Outputs validation evidence for deployment review or audit workflows.

Initial security checks

Shell or code execution enabled without approval

External email or upload tools combined with sensitive data

MCP servers from untrusted sources

Persistent memory without retention rules

Delegation enabled without max depth

High-risk actions missing human approval

Tool allowlist missing

Secrets access available to agent runtime

Planned integrations

LangGraph

MCP

OpenAI Agents SDK

GitHub Actions