Designing Kyro: Local-First Desktop Automation with Tauri v2, Rust, and MCP

1. The Screen Automation Challenge

While terminal agents like M31A solve code repository mutations, developers and knowledge workers spend most of their time in visual applications: web browsers, Figma, design tools, chat clients, and proprietary GUI dashboards.

Recent attempts at "computer use" agents have predominantly been cloud-wrapped Python scripts that capture screenshots every 2 seconds and stream raw bitmaps to remote third-party API servers.

This cloud-dependent paradigm introduces severe obstacles:

1. Unacceptable Latency: Uploading 4K PNG frames over residential internet creates a 2.5–5 second delay per mouse click or keystroke.
2. Severe Privacy Breaches: Screen captures inadvertently expose personal banking tabs, confidential client emails, private Slack messages, and authentication cookies to cloud logging endpoints.
3. Electron Memory Bloat: Typical desktop AI apps bundle full Chromium runtimes that consume 800MB–1.5GB of RAM before performing any work.

To solve this, we engineered Kyro (kyro.tonmoyinfrastructure.org).

---

2. Why Tauri v2 and Rust?

Rather than packaging another heavy Electron shell, Kyro uses Tauri v2 with a native Rust backend and a lightweight React 19 frontend:

Metric	Electron Wrapper	Kyro (Tauri v2 + Rust)	Improvement
Idle Memory (RAM)	680 MB – 1.1 GB	< 95 MB	~85% less memory
Binary Bundle Size	185 MB – 240 MB	38 MB	80% smaller
Local Frame Processing	Web APIs (slow)	Native X11 / Wayland / Quartz	Microsecond capture
System Secret Storage	Plaintext config files	OS Keyring Vault	Hardware-backed security

+--------------------------------------------------------------+
|                     REACT 19 FRONTEND                        |
|        Interactive visual timeline & approval cockpit        |
+--------------------------------------------------------------+
                               |
                               | Typed Tauri IPC (zero-copy)
                               v
+--------------------------------------------------------------+
|                    KYRO RUST NATIVE CORE                     |
|  - Screen Frame Capture (x11rb / wayland-client / Quartz)    |
|  - Input Simulation (enigo / native evdev)                   |
|  - Model Context Protocol (MCP) JSON-RPC Engine              |
+--------------------------------------------------------------+
                               |
                               +---> Local SQLite (WAL Mode)
                               +---> OS Keyring (Vault)

---

3. The Zero-Trust Security Architecture

Kyro enforces a strict zero-trust security model:

1. Local-Only Persistence: All conversations, task histories, action logs, and vector embeddings are stored in a local SQLite database running in WAL mode on your disk (~/.local/share/kyro/data.db). Zero telemetry leaves the machine.
2. OS Keyring Vault (Vault): Model provider API keys (OpenAI, Anthropic, OpenRouter, Ollama) are never stored in plaintext JSON or .env files. They are encrypted using the host operating system's native credential manager (Linux Secret Service, macOS Keychain, Windows Credential Manager).
3. Argument-Aware Authorization: Dangerous operations (such as deleting files, submitting forms with payment data, or sending external emails) trigger an explicit visual confirmation prompt in the Kyro desktop interface.

---

4. Model Context Protocol (MCP) Integration

Kyro includes a native JSON-RPC client for Anthropic's Model Context Protocol (MCP). This enables Kyro to discover and execute tools exposed by local developer environments:

• Reading files from local git repositories.
• Querying local Postgres or SQLite databases.
• Triggering local test harnesses and reporting structured results.

By combining visual UI automation with protocol-driven MCP tools, Kyro acts as a unified autonomous copilot for the entire desktop operating system.

---

5. Experience Kyro

Kyro is open source. You can download pre-built Linux AppImage binaries, macOS DMG packages, or build from source:

Visit the official documentation portal at kyro.tonmoyinfrastructure.org or inspect the repository at github.com/eshanized/Kyro.