Eyes and hands for AI agents on the desktop. A Model Context Protocol (MCP) server for screen perception and input control.
Ghost is a Rust MCP server (~30 tools) that gives an agent the ability to see the screen (capture, OCR, UI-element detection) and act on it (click, type, scroll, drag), plus a record→replay recipe engine. Windows-first. It's a component of Ryu but runs standalone with any MCP-capable client.
- 👁️ See. Screen capture, OCR, and UI-element detection over the accessibility tree.
- ✋ Act. Click, type, scroll, and drag through synthetic keyboard/mouse input.
- 🎬 Record → replay. Capture a task once, replay it deterministically as a parameterized recipe.
- 🧩 ~30 MCP tools. Drop into any MCP client, or let Ryu Core spawn it as a sidecar. Stdio, no network port.
- 🪟 Windows-first. Cross-platform (macOS/Linux) perception + input backends are in progress.
- 🔓 Open & auditable. Dual-use by nature, so the behaviour is open source and, inside Ryu, consent-gated.
| Path | What |
|---|---|
apps/ghost |
the MCP server binary |
crates/ghost-core |
core automation primitives (recipes, store) |
crates/ghost-eyes |
screen perception / vision |
crates/ghost-hands |
synthetic keyboard/mouse input |
cd apps/ghost && cargo build --release # → a stdio MCP server (no network port)Launch it from any MCP client, or let Ryu Core spawn it as a
sidecar. Config + cache live under ~/.ghost/.
Screen perception + synthetic input control are exactly what malware wants. Ghost is open-source so the behaviour is auditable, and inside Ryu it runs only behind explicit user consent. If you embed it, gate it behind clear consent and treat it as a high-trust dependency. See SECURITY.
Ghost is derived from Ghost OS by Ghostwright (MIT). The original copyright + license notice are retained in NOTICE. Ghost is licensed under Apache-2.0 (see LICENSE) with MIT-licensed portions per NOTICE. © 2026 A Major Pte. Ltd.