This post, written in Chinese by Phodal Huang, announces that AutoDev Remote Agent — part of AutoDev
Workbench and one component of the next stage of AutoDev (unit-mesh/auto-dev) — had entered a
trial phase. The post lists what it can be used for: running as an MCP service on a server, helping analyse
GitHub issues and plan tasks, being integrated into a development process, and writing code, designing
architecture and writing test cases, with automated coding, testing and deployment (limited to GitHub Pages)
named as future goals. Its code is published in the unit-mesh/autodev-workbench GitHub repository, and readers
are invited to test it by opening issues there.
The title’s question — why let AI analyse requirements and design solutions only during the day — frames remote agents as assistants that complete development tasks in the background, a direction the author says tools such as Claude Code and the Codex agent are also pursuing.
Key Points
- The author explains why AutoDev is not building its own IDE: from earlier work building an IDE on IntelliJ Community he concluded that an individual cannot bear the machine and cloud-server costs of an independent IDE, and he says he could not accept VS Code’s syntax analysis and refactoring capabilities.
- He characterises mainstream AI coding in 2024 as an RPC/Go-plus-Rust agent backend, a WebView front end for cross-platform IDEs, and vector-based code indexing and retrieval, and argues that in 2025, with Grep/Ripgrep search becoming the preferred alternative to vectorization — which he says is no longer important — and coding models such as Claude able to code agentically, agents can run on the server to write, test and deploy code.
- The prototype, built over the preceding month or more, can be used in a GitHub project’s Actions to analyse issues, plan tasks and write code; the post links it to a GitHub Marketplace action.
- The author says the first version of the agent’s design was produced with the help of Augment, which he calls the strongest AI coding assistant to date, building on a core refactored out of AutoDev’s VS Code version and on AutoDev Workbench’s assistant design.
- The agent’s tool design follows AutoDev Sketch’s, adding GitHub tools on top of MCP-wrapped general tools so that it can fetch issues and write analysis and planning results back to them; an example run shows 18 tools loaded with DeepSeek as the model provider.
- A “Round” mechanism, which the author says Augment created during the design, limits the number of rounds of conversation to avoid infinite loops, with each round making tool calls and summarising them, ending in a complete task plan.
- The post compares the agent’s tools with those of Claude Code, Augment, Cursor Agent and Codex Agent across file operations, terminal execution, process management, code search, code analysis, GitHub integration, network functions, Jupyter support, memory management and visualisation; in that comparison only AutoDev Remote Agent is marked as having GitHub integration. The author acknowledges that the tools still have many bugs.
- For safety when running in GitHub Actions, the author describes sandboxing by creating a complete code-running environment inside a GitHub Action and by dynamically creating new GitHub Actions; he says Docker was set aside for now because it performed poorly on his old MacBook.
- Next-stage goals named in the post are making AutoDev Remote Agent bootstrap itself and adding process-related tools.
Context
The post is an announcement of the author’s own open-source project, and its tool comparison is the author’s own table rather than an independent evaluation.