code-cli

Autohand Code CLI is a fast, self-improving AI coding agent that runs natively in the terminal. It is designed for the environment where developers already work, able to read project context, plan changes, edit files, run tools, and seek user approval before performing risky operations.

[Purpose] Autohand Code CLI is used to plan, reflect, remember, edit, test, and automate work in a codebase. After installation, run autohand and describe the desired result in natural language. It can be used locally, with an editor, or in CI/CD to automate repetitive engineering work, while keeping the user in control of the process.

[Key Capabilities]

• Terminal-native agent: understands the codebase and executes approved changes from the CLI.
• Planning and tools: combines reasoning, file editing, Shell commands, and web context in one loop.
• Interactive REPL: provides a clean terminal experience with file mentions, slash commands, and keyboard shortcuts.
• Modular skills: extends workflows through specialized instruction packages.
• Multi-provider support: works with OpenRouter, LLMGateway, OpenAI, AWS Bedrock, DeepSeek, Azure Foundry Models, Z.ai, and local models.
• Git integration: full version control support, including automatic commits.
• Cross-platform: supports macOS, Linux, and Windows.
• Built-in 40+ tools, covering file operations (read_file, write_file, append_file, apply_patch, search, search_replace, semantic_search, list_tree, etc.), Git operations (git_status, git_diff, git_commit, git_add, auto_commit, etc.), command and dependency management (run_command, custom_command, add_dependency, remove_dependency), planning and memory (plan, todo_write, save_memory, recall_memory), tool registration and search, meta-tool creation, Jupyter notebook cell editing, and team and collaboration-related tools.
• Extension system: canStrictly declarative manifests package reusable tools and agents that can be verified and installed without changing CLI source code; checked and installed runtime extensions with --trust can also register slash commands, Ink UI, status/help snippets, keybindings, CLI flags, hooks, providers, and permission policies.
• Goal management: persistent goal data is retained in the workspace across sessions; supports /goals view, resuming paused goals, queued goals, and more.
• Mobile pairing: works with the iOS app for permission/change request sync, read-only viewing of GitHub delivery metadata, merging ready PRs from the phone, and more.
• Trace monitoring: comes with the ahtraces local trace monitoring tool, supports enabling local monitoring and metadata sync via commands such as autohand --traces-on, and uploaded traces can be viewed and deleted in the console.
• Computer use: the installer installs Autohand Computer Use, built independently on the MIT-licensed Cua engine and connected as a runtime MCP server; built-in computer control skills can be activated via instructions such as "open my browser", and autohand computer status/install/doctor commands are provided to check permissions.
• SDK: provides beta packages for multiple languages including TypeScript, Go, Python, Java, Swift, and more, allowing developers to build capabilities based on the same CLI agent runtime in their own tools, services, workflows, or editor integrations.

[How to Use] The recommended installation method is the fully qualified command, which installs and trusts only the Autohand formula. All supported installation methods expose the same CLI, with command names autohand, autohand-code, agent, and the short name ah, where autohand is the canonical name. After installation, run autohand to enter interactive mode: type / to trigger slash command suggestions, @ to trigger file autocomplete, $ to trigger skill autocomplete, ! to run terminal commands, ESC to cancel an in-progress request, Ctrl+C twice to exit, Shift+Tab to cycle through edit, plan, YOLO, and auto modes, and ? to toggle the keyboard shortcuts panel. It also supports entering the full REPL with no arguments, or using --output-format stream-json, --json stream, etc. to run a single instruction non-interactively in CI/CD,scripts, or quick tasks. You can open the goals view via /goals and toggle it open or closed with Ctrl+G (Cmd+G on macOS).

[Use Cases]

• Collaborate with AI locally or in an editor to understand, modify, test code, and perform Git operations.
• Automate repetitive engineering tasks in CI/CD.
• Build custom workflows or editor integrations through extensions, skills, and the SDK.
• Team collaboration scenarios using built-in team and task tools.
• Combine with the companion iOS app for mobile remote control and PR merging.

[Limitations]

• Risky operations require approval by default unless a different permission mode is selected.
• When command output is large, only the beginning and end are retained; a single output stream is capped at 1,048,576 characters, and the excess is shown with an ellipsis marker; when complete results are needed, output should be redirected to a file.
• agent is a generic name, and the installer automatically scans all writable directories in PATH and replaces an existing agent command; if another tool's agent command stops working, this may be the reason.
• Real paths uploaded from mobile are restricted to prevent symlink escape, with a maximum of 12 files per upload and no more than 15 MB per file.
• Specific fees, prices, platform support details, and effects are not provided in the materials, so no relevant information is available at this time.