charmbracelet/crush
Crush
Crush is a framework designed to orchestrate and secure the execution of external tools invoked by large language models. It functions as a middleware layer that manages the flow of agentic tool calls, providing a controlled environment for terminal-based automation and task processing.
The project distinguishes itself by implementing a policy-driven security layer that intercepts, validates, and modifies tool execution requests. By wrapping command calls within a process-boundary layer, it allows for the automated approval of specific operations and the dynamic injection of contextual metadata into the execution environment.
Beyond its core orchestration capabilities, the tool supports the processing of structured data directly within the command line. This enables the transformation of complex inputs into simplified formats, facilitating the creation of automated data pipelines and secure terminal workflows.
Features
- Agentic Tool Orchestration - Orchestrates and secures the execution of external tools invoked by large language models.
- AI Agent Orchestration Frameworks - Acts as a framework for managing and automating the execution flow of tools invoked by language models.
- Tool Execution Interceptors - Intercepts, validates, and modifies tool calls made by AI agents to enforce security policies.
- Terminal Workflow Automations - Provides a security layer for command-line tools to prevent unauthorized actions during automated execution.
- Command Interception Hooks - Intercepts and validates tool execution requests before they reach the system shell.
- Coding Agents - Terminal-based AI coding agent.
- Coding Agents and IDEs - Agentic coding tool with LSP integration and terminal UI.
- Command Line Tools - Multi-model coding agent featuring a beautiful terminal user interface.
- Entwicklungstools - AI-powered coding agent.
- Command Line Automation Tools - Streamlines terminal tasks by automating permissions and injecting metadata into command environments.
- Policy-Driven Execution Guardrails - Implements policy-driven security rules to automate approval workflows for tool execution.
- Task Automation Tools - Automates tool permissions to bypass manual confirmation prompts during routine operations.
- Process Execution Wrappers - Wraps external tool calls in a controlled execution layer to monitor and manipulate command inputs and outputs.
- Execution Message Injection - Appends instructions and metadata to the execution environment to guide agent behavior.
- Environment Variable Managers - Injects contextual metadata into the execution environment to guide agent behavior during tool invocations.
