An AI agent that can execute tasks using tools and manage conversations.
- Interactive and single-run modes
- Streaming text responses
- Multi-turn conversations with tool calling
- Configurable model settings and temperature
- File operations: read, write, edit files
- Directory operations: list directories, search with glob patterns
- Text search: grep for pattern matching
- Shell execution: run shell commands
- Web access: search and fetch web content
- Memory: store and retrieve information
- Todo: manage task lists
- Automatic context compression when approaching token limits
- Tool output pruning to manage context size
- Token usage tracking
- Multiple approval policies: on-request, auto, never, yolo
- Dangerous command detection and blocking
- Path-based safety checks
- User confirmation prompts for mutating operations
- Save and resume sessions
- Create checkpoints
- Persistent session storage
- Connect to Model Context Protocol servers
- Use tools from MCP servers
- Support for stdio and HTTP/SSE transports
- Specialized subagents for specific tasks
- Built-in subagents: codebase investigator, code reviewer
- Configurable subagent definitions with custom tools and limits
- Detects repeating actions
- Prevents infinite loops in agent execution
- Execute scripts before/after agent runs
- Execute scripts before/after tool calls
- Error handling hooks
- Custom commands and scripts
- Configurable working directory
- Tool allowlisting
- Developer and user instructions
- Shell environment policies
- MCP server configuration
- Terminal UI with formatted output
- Command interface: /help, /config, /tools, /mcp, /stats, /save, /resume, /checkpoint, /restore
- Real-time tool call visualization