AutoDev is an AI coding tool whose releases are published in the unit-mesh/auto-dev GitHub repository. This page covers the tool in that repository; AutoDev is a separate page for a
different framework that shares the name. As described
in AI 代码审查再进化：AutoDev 多智能体协作架构深度解析, it can be installed as a CLI from the
npm package @autodev/cli or downloaded as AutoDev Desktop.
The problem its code-review feature addresses, as its author presents it, is that the information a reviewer needs — lint results, tests, issues, change history — is scattered across systems, no single tool understands complex logic, and manual review is slow and subjective.
Capabilities
IDE plugin
Its author presents AutoDev as a code-change assistant: rather than regenerating code for each new requirement, it helps developers change existing code, and he argues this requires building conventions, best practices and software knowledge engineering into the tool. Version 1.8 of the plugin, described in 构建演进式 AI 辅助编码，融合 DevOps 规范和实践, added several features along those lines: commit messages generated from the code change together with the user’s current requirement ID, fetched through an internal OA system; refactoring driven by the code smells the IDE’s own inspections report; five AI name suggestions offered when the user invokes the IDE’s rename feature (enabled manually in the settings); and terminal command generation that puts the date, operating system and shell into the context. The same version added Chinese settings pages and prompts, an easier LLM server test, support for 2024.1 IDE versions and AutoSQL improvements.
Custom agents and DevIns
An earlier AutoDev release added custom agents, letting users build their own agents to assist with software
development tasks. Its author describes AutoDev as assisting development through context awareness and
custom capabilities — automatic test generation and running, and UI generation among them — and notes that
its custom prompts can refer to the editor state, as in a prompt asking the AI to explain the selected code.
A following release, announced in AutoDev DevIns —— 开源 AI 智能体语言，构建 AI 驱动的自动编程, added
DevIns, a language between natural language and instruction text: with version 1.7.2 of the plugin installed, a user writes a
.devins file and runs it, AutoDev compiles its instructions together with the surrounding context into
instruction text for the agent, and instructions such as /write, /run, /patch and /commit let the
result be written to files, tested, patched and committed.
Code review
In version 0.3.0 of the CLI, review is run with autodev review -p .. The review proceeds as a four-step
pipeline: it collects static information (changed hunks from the Git diff, affected classes and methods located
with tools such as CodeGraph, results from linters such as ESLint, Ktlint and Detekt, and related issues and
tests); has an LLM analyse it along a chosen review type (comprehensive, performance, security or style) to
produce structured findings; merges and prioritizes lint and AI findings into a fix plan that the user can edit;
and generates fixes as a patch that can be rolled back or iterated on.
The work is split across agents. A main CodeReviewAgent aggregates information and orchestrates; analysis
sub-agents (AnalysisAgent, ErrorRecoveryAgent, CodebaseInvestigatorAgent) handle large content, error recovery
and repository-wide investigation; and a CodingAgent makes the code changes through tool calls such as
read_file, write_file, lint and test. Sub-agents are managed by a SubAgentManager, and all tools are
registered in a ToolRegistry and executed through a ToolOrchestrator.
CLI and agent test framework
The CLI is built on AutoDev’s MPP (Multiplatform Paradigm) core. Its author explains in
AutoDev CLI：构建 AI Agent 生成的 AI Agent 质量保障与验证架构 that it
was designed to fix a testing problem in the earlier IntelliJ-based AutoDev: agent logic was bound to IDE UI
events and IDE kernel objects and could not easily run in CI/CD. Because the same CodingAgent code and renderer
are shared across the CLI, desktop and mobile targets, he treats the CLI as the place to reproduce and fix agent
problems that would very probably appear on the other targets too. Invoked as autodev code --path . --task "...", it logs tool registration, sub-agents and MCP servers, and those logs become the input an AI agent uses
to fix the next problem.
On top of it, the author had AI generate a dedicated test framework for the coding agent that analyses three things per scenario — the prompts generated, the tools called and the code changes produced — with scenarios stating expectations such as a minimum number of file reads and writes or particular files being created or modified. The framework runs in a GitHub Actions workflow with a pass threshold, which the author asked to be at least 80%.
AutoDev Remote Agent
AutoDev Remote Agent, part of AutoDev Workbench and published in the unit-mesh/autodev-workbench repository,
entered a trial phase in June 2025, as announced in AutoDev Remote 编程智能体：你何必只让 AI 在白天分析需求、设计方案. It can run as
an MCP service on a server or inside a GitHub project’s Actions, where it analyses GitHub issues, plans tasks and
writes code, writing its analysis and plans back to the issue. Its tool design follows AutoDev Sketch’s, adding
GitHub tools on top of MCP-wrapped general tools, and a round limit keeps its conversations from looping
indefinitely. Its sandboxing relies on creating a complete code-running environment inside a GitHub Action and
on dynamically creating new GitHub Actions.
Adoption & Ecosystem
Its author has said that AutoDev will not focus on building its own IDE, citing the cost to an individual of maintaining one, and that the remote agent reflects a view that, as coding models became able to code agentically, agents could run on the server to write, test and deploy code. The Remote Agent’s first version was designed with the help of the Augment coding assistant, building on a core refactored out of AutoDev’s VS Code version; making it bootstrap itself is a stated next goal.
For code review, its author says future versions will integrate test coverage, CI/CD and incremental analysis.