Computer Use Agent

A TypeScript CLI agent that translates natural language into shell commands and executes them autonomously. Supports multiple LLM providers, trust levels, file operations, and undo.

Features

• Natural Language to Shell: Describe what you want, the agent figures out the commands
• Agentic Loop: Executes multi-step tasks autonomously (up to 25 steps)
• 5 Built-in Tools: bash execution, read_file, write_file, list_directory, search_files
• Streaming Responses: See the agent's thinking in real-time
• 4 Trust Levels: Sandbox, Standard, Trusted, Unrestricted — control what the agent can do
• Multi-Provider: OpenRouter (free), OpenAI, Anthropic, Ollama (local)
• Project Detection: Auto-detects Node.js/Python projects and loads context
• Undo System: Git-based snapshots let you revert any change
• Session Persistence: Save and resume sessions

Setup

# Install dependencies
npm install

# Configure API key
cp .env.example .env
# Edit .env and add at least one provider key

# Build
npm run build

Providers

Provider	Key	Get Key
OpenRouter	OPENROUTER_API_KEY	https://openrouter.ai/keys (free)
OpenAI	OPENAI_API_KEY	https://platform.openai.com/api-keys
Anthropic	ANTHROPIC_API_KEY	https://console.anthropic.com/
Ollama	No key needed	https://ollama.ai (local)

Usage

# Start (prompts for provider, model, trust level)
npm start

# Auto-execute mode (no confirmation prompts)
npm run start:auto

# Development mode
npm run dev
npm run dev:auto

# With specific model and trust level
npm start -- -m gpt-4o -t trusted

# Combine flags
npm start -- -m gpt-4o -t trusted -a

Trust Levels

Level	Allowed	Use Case
Sandbox	Read-only: ls, cat, grep, find	Exploring repos safely
Standard	+ file creation, mkdir, tools	Daily development (default)
Trusted	+ rm, mv, pip/npm install	Full development
Unrestricted	Everything except sudo, shutdown	Power users

Change trust level at any time with the trust command.

Commands

• quit / exit / q — Exit
• clear — Clear screen
• cwd — Show current directory
• trust — Change trust level
• help — Show help
• undo — Revert last change (requires git)
• diff — Show changes since last snapshot
• snapshots — List session snapshots
• save [name] — Save current session
• sessions — List saved sessions

CLI Flags

• -v, --verbose — Enable debug output
• -m, --model <model> — Override LLM model
• -t, --trust <tier> — Set trust tier
• -a, --auto — Auto-execute without confirmation
• -k, --api-key <key> — Override API key

Architecture

src/
├── main.ts              # CLI interface, agentic loop
├── llm.ts               # OpenAI-compatible client with streaming
├── config.ts            # Trust tiers, security, provider configs
├── bash.ts              # Command execution
├── messages.ts          # Conversation history
├── types.ts             # Type definitions, Zod schemas
├── utils.ts             # Utility functions
├── project-context.ts   # Auto-detect project type
├── undo-manager.ts      # Git-based undo system
├── session-manager.ts   # Session persistence
├── help.ts              # Help system
└── tools/
    ├── registry.ts          # Tool registry and dispatch
    ├── bash-tool.ts         # Bash execution tool
    ├── read-file-tool.ts    # File reader with offset/limit
    ├── write-file-tool.ts   # File writer with append mode
    ├── list-directory-tool  # Directory browser
    └── search-files-tool.ts # Grep-based search

Security

• Commands validated against trust-tier allowlist
• System-critical commands always blocked (sudo, shutdown, reboot, mkfs, etc.)
• Destructive commands require confirmation even in trusted mode
• Read-only commands skip confirmation
• 30-second timeout per command
• Git snapshots before every mutation

Development

npm run dev          # Run with ts-node
npm run build        # Compile TypeScript
npm run build:watch  # Watch mode
npm run lint         # ESLint
npm run format       # Prettier
npm run clean        # Remove build artifacts

Requirements

• Node.js >= 18
• At least one LLM provider API key (OpenRouter has free models)