AutoDev 3.0 Xiuper: five code-backed agents and a root Gradle graph that skips three modules

Three modules are in the repo but outside the root build graph

The status snapshot is the most concrete part of this README, because it separates what exists from what compiles. The default root Gradle build includes mpp-core, mpp-ui, mpp-codegraph, mpp-server, mpp-viewer, mpp-viewer-web, xiuper-ui, xiuper-fs, and xiuper-e2e, with mpp-idea attached as a composite build. Against that, mpp-vscode, mpp-ios, and mpp-web are maintained in the repository but are not part of the current root settings.gradle.kts default build graph.

So a clone contains the VS Code extension, the iOS shell, and the Vite landing page, and a successful root build touches none of them. The consequence is that a green root build says nothing about the extension or the mobile wrapper, and anyone evaluating the VS Code path has to build that package on its own.

Six delivery surfaces, with iOS left as build from source

Getting hold of the platform means choosing a channel, and the channels are not uniform. The IntelliJ IDEA plugin comes from the JetBrains Marketplace as autodev-experiment, the VS Code extension from the Visual Studio Marketplace as Phodal.autodev, the web app from web.xiuper.com, and desktop and Android builds from GitHub Releases. The CLI is the only npm package:

npm install -g @xiuper/cli

iOS has no artifact at all and is listed as build from source. Alongside all of this, AutoDev 2.0 is still called Stable and lives on the autodev2 branch with its own plugin id. The consequence is that an alpha is being distributed through five channels at once, iOS users compile rather than install, and AutoDev 2.0 lives on a different branch.

Five agents are code-backed, and the web flows are not

The built-in agent table marks status per row, which makes the weak spot easy to locate. DocumentAgent covers document querying and feature tree generation with DocQL and an analysis sub-agent. CodingAgent carries workspace tools, orchestration, file system, shell, MCP, sub-agents, and AGENTS.md. CodeReviewAgent does review, lint summaries, and fix generation across a linter, diff review, and PR review services. ChatDBAgent does natural-language database work with schema linking, SQL generation, and SQL revision. ArtifactAgent produces self-contained runnable outputs in HTML, React, Node.js, Python, SVG, and Mermaid. All five are marked code-backed. The Web Agent and WebEdit rows are the exception, marked experimental, covering page inspection, DOM context, and chat-assisted actions. The consequence is that the web interaction workflow named in the platform's own introduction is the part with no stability claim behind it.

SubAgents are invoked as tools, and their platform support is uneven

SubAgents are described as specialized micro-agents invoked by the main Coding Agent for focused tasks, following what the README calls the agent as tool architecture pattern. The platform support column is where the table earns its keep. The NanoDSL Agent generates AI-native UI code from natural language descriptions, with a token-efficient DSL, component generation, state management, and HTTP requests, and it is marked for all platforms. The Chart Agent targets the ComposeCharts library with pie, line, column, and row charts plus data analysis, also all platforms. The PlotDSL Agent, for statistical charts and data visualizations from natural language, is limited to JVM Desktop and Android, with ggplot2-inspired syntax, multiple chart types, themes, and Lets-Plot rendering. The consequence is that a chart request behaves differently by target, and the restriction sits in a support column rather than in the description.

AGENTS.md is discovered from the project hierarchy and injected

Project rule awareness is automatic: AGENTS.md is discovered and injected into prompts from the project hierarchy, with no configuration step. The Coding Agent's own capability list includes AGENTS.md alongside the file system, shell, MCP, and sub-agents, and the repository carries both AGENTS.md and AGENTS.md.example at its root, so the example file is the template to copy. Around that sits ecosystem support for MCP, A2A agent commands, Claude Skill loading, and SpecKit command integration. The consequence is that agent behavior changes based on files in whatever repository you open, which is the point and also the surprise.

Tree-sitter parsing covers seven languages and six model providers

Two capabilities are worth separating from the agent list because they bound what the platform can read and talk to. Code intelligence is Tree-sitter based, with parsing for Java, Kotlin, Python, JavaScript/TypeScript, Go, Rust, and C#, and the platform is described as fully internationalized in Chinese and English. Multi-LLM support names OpenAI, Anthropic, Google, DeepSeek, and Ollama, and more beyond them. The example directory holds example/custom_llm_server/ alongside example/custom_agent/, example/mcp/, example/prompt/, and example/Hello World.ipynb.

Testing is E2E-shaped and has no top-level agent

One sentence in the agent section is unusually candid. Testing-related capability currently exists mainly through xiuper-e2e and E2E-oriented agent/tooling work, and it is not yet represented as a clearly separated top-level agent in the current code layout. That matters because every other capability area in the table has a named class behind it, so the platform can be asked for DocumentAgent, CodingAgent, CodeReviewAgent, ChatDBAgent, or ArtifactAgent by name. Testing cannot. The consequence is that xiuper-e2e sits in the default build as a shared support module next to xiuper-fs and xiuper-ui, and anyone expecting a testing agent to appear alongside the others in the menu will not find one in this layout.

Alpha tags stopped in February while the default branch kept moving

The release history and the branch history disagree about how current the software is. The newest tags are compose-v3.0.0-alpha5 and compose-v3.0.0-alpha4, both published on 2026-02-07, preceded by compose-v3.0.0-alpha3 on 2025-12-29, and the last push on the default branch master is dated 2026-09-25. That is roughly seven months of commits past the newest artifact. The repository is also candid about its own documentation practice, carrying .adr.json for architecture decisions, CHANGELOG.md, AGENTS.md, a render.yaml, a settings.docker.gradle.kts beside settings.gradle.kts, and gradle-252.properties. Beside the Gradle files the tree also carries commonMain/, src/, prompts/, docs/, example/, examples/, .run/, .idea/, and .augment/, and the LICENSE is MPL-2.0 rather than a permissive plugin license. The consequence is that the newest tagged release is about seven months behind master.

Editorial conclusion

AutoDev 3.0 Xiuper suits an evaluator who wants to see a Kotlin Multiplatform agent platform assembled around one shared core, and a team that already uses AGENTS.md and wants the same rules injected into an IDE, a CLI, and a web app from one codebase. It does not suit anyone expecting a stable 3.0: the README titles it Alpha, the newest tag is compose-v3.0.0-alpha5 from 2026-02-07 while the last push is 2026-09-25, and the stable line is AutoDev 2.0 on a separate branch. Before adopting it, check whether the module you need is in settings.gradle.kts, decide whether the web flows marked experimental are on your critical path, and read the LICENSE, which is MPL-2.0 rather than the permissive license most plugin work assumes.

Frequently asked questions

Official sources

1. License: MPL-2.0
2. phodal/auto-dev on GitHub
3. Project website
4. README
5. Releases