DeepSeek has released Super-agent, an open-source coding agent prototype that monitors a single GitHub repository and can optionally listen to Feishu group chats. The system routes approved requests into separate work sessions, completes assigned tasks and reports progress or results in the conversation where each request originated.

DeepSeek’s Harness team announced the project on October 31 and released the code under the MIT license. The launch drew comparisons with Anthropic’s Claude Code, although DeepSeek describes Super-agent as research software rather than an official product or production-ready tool. The company also said the public version differs from the system used internally.

The project’s code is available on GitHub.

How Super-agent handles GitHub and team requests

The prototype is designed to work with one repository on github.com and can receive requests through GitHub or optional Feishu group chats. An approved member can submit a request through either channel. Super-agent converts the request into a work item, creates a dedicated session and working directory, and returns progress updates or a completed result in the original conversation. The result may include a pull request.

A local web panel displays issues that require a person’s decision. The system uses a persistent Router session to process requests and maintain shared context, while assigning individual tasks to separate Worker sessions. The Router can accept additional requests while existing tasks are running, allowing work to overlap. Workers respond to comments on their pull requests, and each Ticket receives its own session and directory.

This workflow connects chat requests with repository changes and review activity, so deployment decisions involve both software configuration and team governance. The documented approval flow, observation mode and user restrictions provide controls for evaluating those interactions before the agent is allowed to act automatically.

Teams can configure the agent to perform work according to written rules, but the default setup is observation mode. In that mode, Super-agent records what it would do without executing the proactive task. Operators can review the proposed actions before enabling automatic acceptance. DeepSeek says this arrangement allows maintainers to adjust permissions and limit unwanted activity.

The project attracted early attention after its release. GitHub showed 6.6k stars and 652 forks on November 1. Those figures indicate interest in the repository, not coding performance. DeepSeek’s README warns that the public code may contain bugs, uses inconsistent engineering practices and differs from the team’s internal version. It recommends reviewing and adapting the code before deploying it in another environment.

Setup requirements and access risks

Super-agent requires a dedicated GitHub user with write access to the target repository. Its setup supports Node.js 22.19 or later, or version 24 or later, and also requires pnpm, Git 2.32 or later, Python 3.8 or later and tmux. Operators need GitHub CLI 2.48 or later, a supported DeepSeek Harness version and a configured model provider.

The project currently supports repositories hosted on github.com but not GitHub Enterprise Server. Feishu integration is optional and requires a bot and lark-cli. A separate account is used for the agent’s GitHub identity and signing, adding several local tools and account configurations to the installation process.

The more significant consideration is the agent’s local process access. Router and Worker sessions run with the full permissions of the operating-system user who launches them. A Worker can execute programs by path and read that user’s home directory. GitHub credential wrappers select the agent’s GitHub account, but they do not isolate the process from the rest of the operating system.

For that reason, DeepSeek recommends using a dedicated operating-system account containing only the agent’s credentials. A Worker may otherwise access files and credentials available to the launching account, and the agent’s GitHub identity does not restrict its local permissions. Because Super-agent can access team chats and repository activity, operators must also review which users may submit requests and what repository permissions are available. DeepSeek’s documentation indicates that teams should begin with narrow assignments while evaluating the system.

Bloomberg framed the release as a challenge to Anthropic’s Claude Code. Super-agent routes work through chat, source control and code review, but its repository contains no comparative benchmark covering coding accuracy, speed or reliability. It also provides no published evidence that the DeepSeek coding agent matches Claude Code’s coding performance. For teams assessing the prototype, repository-hosting support, local permission isolation and independent performance testing therefore remain practical evaluation points rather than settled capabilities.

Source: Blockonomi