How AutoDev Uses Kotlin Multiplatform & Compose for True Multi‑Device AI Coding

Six months after introducing AutoDev's next‑generation architecture focused on multi‑device support, the author resumed development, completing a proof‑of‑concept that validates new interaction and architectural ideas.

AutoDev Multi‑Platform Refactor Using KMP + Compose

The new architecture centers on an MPP‑Core built with Kotlin Multiplatform, delivering both JavaScript and JVM implementations. These implementations enable a range of products:

JS version: @autodev/cli provides a Command Line Interface (CLI) and Terminal UI (TUI).

JVM version powers AutoDev IDEA, AutoDev Desktop, AutoDev Android, and the planned AutoDev iOS.

To mitigate high AI token costs and the limitations of current agents, the team created a TreeSitter‑based MPP‑CodeGraph that compresses context using tree and graph structures, reducing the need for expensive services such as Gemini CLI or Claude Code.

CodeGraph: Tree‑and‑Graph Based Context Compression Engine

Although Gemini CLI and Claude Code offer near‑unlimited context, they remain unreliable and costly for complex refactoring tasks. CodeGraph addresses these challenges by enabling efficient code search and analysis across platforms, leveraging TreeSitter’s WebAssembly‑compatible parsing capabilities.

AutoDev IDEA: Built‑in MCP Capability

In the new IDEA version, the Multi‑Channel Programming (MCP) capability is embedded directly, while AutoDev itself functions as an MCP server that other agents can invoke.

Why Desktop and Android?

The author questioned whether a new AI‑driven IDE needed syntax highlighting or code navigation, noting that recent tools like Cursor demonstrate the value of an agent‑centric UI. With a lightweight UI, developers can perform Vibe Coding remotely on tablets or phones, freeing them from a constant desktop presence.

Rebuilding the UI with Compose

After experimenting with IntelliJ Compose to build a new AutoDev IDE UI, the author migrated to the 2025.2 IDEA version, finding Compose far more stable and reliable than Swing, and better suited for a chat‑style interface with live preview capabilities.

MPP‑Core: The Cross‑Platform Engine

Compared to last year’s architecture, which varied between JetBrains and VSCode plugins, the current design unifies everything around Tool and Agent abstractions. This separation isolates platform differences, allowing all capabilities to reside in the core and enabling true cross‑platform functionality.

Images illustrating the new architecture, CodeGraph, and UI prototypes are included in the original article.

Developers interested in AI‑assisted programming, cross‑platform agents, or the AutoDev architecture are invited to explore the open‑source repository at https://github.com/unit-mesh/auto-dev, submit feedback, pull requests, or new ideas to help advance the multi‑device coding experience.