manus

Claude's hand on your Mac — a computer-use agent written in Almide.

manus (Latin: hand) gives Claude eyes and hands on a real macOS desktop. It screenshots the screen, sends it to the Anthropic Messages API computer tool, and executes the actions Claude returns (clicks, typing, key presses) with cliclick. Screenshot → think → act, in a loop, until the task is done.

No Docker, no sandbox VM — it drives your machine. That is powerful and risky; every action asks for confirmation unless you opt out.

Requirements

• macOS (Apple Silicon or Intel)
• ANTHROPIC_API_KEY in the environment
• cliclick on PATH — brew install cliclick
• Screen Recording + Accessibility permission granted to the terminal you run it from (System Settings → Privacy & Security)

Build & run

almide build src/main.almd -o manus     # compile to a native binary
export ANTHROPIC_API_KEY=sk-ant-...
./manus "open Safari and search for the Almide language"

Or run straight from source:

almide run src/main.almd -- "take a screenshot and describe what you see"

Environment

Var	Meaning
ANTHROPIC_API_KEY	required — your Anthropic API key
MANUS_MODEL	model id (default claude-sonnet-4-6)
MANUS_YOLO=1	skip the per-action confirmation prompt
MANUS_TASK	task text, if not passed as an argument

Safety

manus controls your real machine. Mis-clicks and mistyped text can do real damage (delete files, send messages). Defaults are conservative:

• Every action prompts [y/N] before it runs. MANUS_YOLO=1 disables this.
• Watch it work — don't walk away.
• Prefer running it in a dedicated macOS user account for anything risky.
• There is a hard cap of 40 steps per run.

How it works

1. Read the logical screen size via osascript (Finder desktop bounds).
2. screencapture the screen, downscale to logical points with sips (so Claude's coordinates map 1:1 onto cliclick's), base64-encode.
3. Call POST /v1/messages with the computer_20250124 tool and the computer-use-2025-01-24 beta header (built directly with the http and json stdlib modules).
4. Parse the returned tool_use block, execute the action via cliclick, and reply with a fresh screenshot as a tool_result.
5. Loop until Claude stops without requesting a tool.

Supported actions: screenshot, left_click, right_click, double_click, middle_click, mouse_move, type, key, wait, cursor_position. (scroll and left_click_drag return a "not supported" result in v0.1.)

Status

Early (v0.1). Written to dogfood Almide on a real, effect-heavy program: subprocess control, HTTP, JSON, an agent loop. Known upstream issue found while building it: almide test (WASM target) trips a Perceus reference-counting verification bug in the compiler on this module, though almide build (Rust target) is clean — see CLAUDE.md.

License

MIT