anumati

anumati is a deterministic auto-approver designed for AI coding agents, aimed at eliminating repetitive approval prompts for safe shell commands. Leveraging a simple configuration of named matchers, anumati automatically permits safe commands such as git status, npx tsc, cargo test, and more, while still prompting for genuinely risky commands. This agent-agnostic solution works seamlessly with both Claude Code using the PreToolUse hook and OpenAI Codex through its PermissionRequest hook.

Key Features

• Deterministic Approval Logic: Command approval is determined through explicit matchers with a strict grammar, ensuring predictability in approvals.
• Safety by Design: Only read, build, and test commands are allowed, while potentially harmful actions like file writes or deletions prompt the user for approval.
• Comprehensive Command Line Handling: Commands containing multiple sub-commands are approved only if each is verified safe, preventing the chaining of dangerous commands under a safe command.
• Self-Extending Configuration: The tool generates suggestions for new matchers whenever a command is not approved, enabling real-time config growth based on user behavior.
• Bash-Centric Focus: Specifically engineered to handle Bash commands, keeping read, write, or edit actions within the permissions of the executing agent.

Workflow

When a shell command is executed, anumati evaluates it against the defined safety rules:

1. If a rule is matched, the command is auto-approved without further prompts.
2. If not, a standard permission prompt is shown along with a suggestion to allow the command on the next attempt.

The command evaluation ensures that undesirable sub-commands cannot bypass checks through command chaining, maintaining a high standard of safety throughout.

Getting Started

To initiate, simply run:

anumati init

This command will automatically prompt for configurations, write starter rules for common commands, and scaffold an audit log for tracking approvals. Users can gradually build their command approval settings as needed, facilitating a more efficient workflow over time.

For more detailed guidance, visit the documentation:

• Configuration Reference
• Multi-Agent Support
• Command-Style Guide

With anumati, driver safety and streamlined command execution are at the forefront, providing peace of mind while leveraging AI coding capabilities.