Qoder

Qoder is an AI IDE published by Alibaba. It is one of the five tools Rule Taxonomy and Evolution in AI IDEs: A Mining and Survey Study selected for study on the basis that each lets developers explicitly define AI IDE Rules the IDE must follow during code generation and chat interactions. That study records its release date as 21 August 2025, per the tool’s official changelog.

Capabilities

Qoder’s rule files live under .qoder/rules/ in the project, and rules are honoured during generation and chat — see AI IDE Rules for the mechanism as it works across tools.

A product Alibaba Cloud documents as Qoder CN

Alibaba Cloud publishes a user guide, on help.aliyun.com under a /zh/lingma/ path, for a product its body text calls Qoder CN. Whether Qoder CN is this product under another name, a regional edition of it, or a different product altogether is not something that guide or the study above states, and nothing else held here settles it. The guide is recorded on this page because the name is the nearest thing either source offers to a link between them; the account below is what that guide says about Qoder CN, and should not be read as established of Qoder.

That guide describes an agent mode (智能体模式) which it says can decide for itself what to do, sense its environment and use tools, carrying a coding task through end to end with project search, file editing and the terminal. It lists four core capabilities. The guide says the agent makes project-level changes, breaking a task down itself and editing several files across a project, with iteration over successive turns and snapshot rollback available. It says the agent senses the project automatically — framework, technology stack, the files a task needs, error messages — so that context need not be attached by hand. It says the agent draws on more than ten built-in programming tools at its own discretion, among them file search, file and directory reading, semantic symbol search within the project, file modification, error retrieval and terminal execution, and that it does so without asking the developer to confirm or intervene. And it says the agent decides on, writes and runs terminal commands as part of carrying out a task.

A separate page of the same guide enumerates those tools by name, grouped by what they are for. The inventory it gives is:

• Retrieval — search_codebase (semantic search over the current project), search_file, grep_code, search_symbol, list_dir, web_search (stated to need no API key), fetch_content (retrieving the content at a URL), search_memory and fetch_rules.
• File editing — edit_file, read_file, delete_file and create_file, which the guide says the agent uses to make changes across several files in a project.
• Terminal execution — run_in_terminal and get_terminal_output, with commands written from what has happened so far in the task.
• Problem retrieval — get_problems, which reads the code problems shown in the IDE’s Problems pane.
• Memory — update_memory, invoked either when a developer asks for something to be remembered or when the tool judges that something should be retained automatically.
• To-do planning — add_tasks and update_tasks, which the guide ties to the agent planning for itself: it breaks pending work into to-do items and then revises the execution path as it goes.

That page also states that MCP services are supported and freely configurable by the developer, which is consistent with the agent page’s account of MCP tools below. Two small inconsistencies in the table are worth noting rather than smoothing over: the heading “终端执行” (terminal execution) appears twice, the second time over the to-do planning rows, and the update_tasks row has its description in the name column.

On confirmation the guide describes two gates. Terminal commands, it says, are by default presented for confirmation before each run, with Run sending the command to the IDE’s terminal window and Cancel skipping it and returning to the agent, which then plans its next step from that feedback; commands meant to run in the background are marked as such, and the agent carries on with other work, checking their output when it needs to. An allowlist of commands that may run without confirmation is described as configurable in the plugin’s chat settings, under an Auto-Run section. Where MCP tools are configured, the guide says the agent decides for itself whether calling one would help, asks before each call, and feeds the result back in as context.

The guide also describes planning and task tracking. For a task it judges complex the agent is said to produce a plan, shown to the developer for review before execution begins, with /plan triggering the same thing deliberately; and to derive a to-do list from the request, extend it when the request is extended, and show each item’s state in the chat window. A separate prompt-optimization control is described as rewriting a rough instruction into a more detailed one using the surrounding context and conversation history, which the developer can then edit, submit or undo.

A further page of the same guide describes a Code Review agent built into agent mode, which it says needs no separate configuration. The developer switches the chat panel to agent mode and uses a /code-review command, or describes the review wanted in natural language, naming the scope: the whole project, specific files, a Git diff or a pull request. The agent is said to review that scope across several dimensions — logic defects, security vulnerabilities (SQL injection and XSS are the examples given), performance bottlenecks and code-style problems — and to produce a structured report sorted by severity into errors, warnings and suggestions, each giving the code location, a description of the problem and a suggested fix. For large changes the page advises reviewing module by module to get finer-grained feedback, and, where a change involves specific business logic, briefly explaining that background in the request. See Code Review Agent for the pattern across tools. That page also carries a notice that a new documentation site for the Qoder CN series has gone live at docs.qoder.cn, and that the content on the guide’s own site may not reflect the product’s latest features and changes.

Note that the guide describes what is being configured as a plugin, with commands sent to a host IDE’s terminal window — which is a different shape from the standalone AI IDE the study above describes Qoder as being. That is a further reason not to treat the two accounts as describing one product.

A third account, in a spec-driven-development survey

A chapter of Jimmy Song’s online handbook 智能体构建指南, surveying representative implementations of Spec-driven development, lists a Qoder among them and describes it as an AI programming assistant designed specifically for SDD scenarios, built on the idea of specification as code. On that account it supports writing specifications in structured Markdown, generates project structure, code and test cases from them, and helps the developer refine and evolve a specification over successive rounds of conversation; it is said to integrate LLM, code generation, testing and deployment capabilities, and to suit team collaboration and complex engineering work.

That chapter gives no URL, vendor or version for the tool it is describing, so whether it means the AI IDE above, the plugin Alibaba Cloud documents as Qoder CN, or something else again is not established by anything held here. It is recorded as a third account under the same name rather than merged into either of the two above.

Adoption & Ecosystem

Two figures from that study bear on adoption, and they measure different things. In its repository mining, an initial search returned 125 candidate Qoder projects; after keyword and rule-file filtering and manual inspection, 1 remained in the final dataset of 83 projects declared as built with an AI IDE. In its practitioner survey, 16 of 99 respondents reported currently using Qoder for development. The mining figure reflects how many public projects both use the tool and say so in their README or description, which the study notes undercounts projects that use an AI IDE without declaring it; the survey figure reflects self-reported use among developers who had committed changes to rule files.