DeepSeek open-sources a free AI coding assistant that works on its own

China's DeepSeek has released a new AI system called dsh-superagent — a programming assistant that works on its own — for free. The system was published on October 31, 2026, as an open-source repository on GitHub, under the MIT licence, which lets anyone download, read and modify it.

Unlike ordinary AI tools that wait for you to ask a question, this assistant is "proactive": it listens on GitHub — and, if a team wants, on the workplace chat app Feishu — and when a person the team has authorised asks for work, it opens a work item, does it in its own session and directory, and then answers in the original conversation, for example with a pull request.

Inside it are two roles: a small main session that routes incoming requests (the Router), and a separate worker session (the Worker) for each accepted job. Its core principle, as DeepSeek writes it, is to take as little human attention as possible while giving as much value as possible: it acts only when asked, or when a written rule from the team allows it, and work nobody asked for starts in a mode that only records what the assistant would have done. It speaks only when a person must act or can use a result.

But DeepSeek itself puts a big warning up front: this is a "research prototype for reference only, not an official DeepSeek product". The published code is not identical to the company's internal system, nor to the one described in its design article, and some ideas in that article are not yet implemented. The company says the code grew out of its team's day-to-day work, with much of the engineering done by coding agents themselves.

Another warning is about security: the deployment guide says both of the assistant's presets run with "danger-full-access" as the operator's operating-system user, and a worker can execute programs by path and read the operator's home directory. The guide therefore advises using a dedicated operating-system account containing only the agent's credentials.

To run it you need a separate GitHub account for the agent, a model provider configured for DeepSeek Harness — DeepSeek's open system for running agents, which is itself in developer preview and can change — and, for chat requests, a Feishu or Lark bot. The code itself is free, but you pay for whatever model access you connect.

The release has drawn fast attention: by 06:13 GMT on Sunday, November 1, less than 21 hours after it was created, the repository had 3,825 stars and 359 forks on GitHub, according to GitHub API figures reported.

What does this mean for you?

For software developers and technology students in East Africa, the biggest value of this release is not running this system straight away on important work — DeepSeek itself says not to treat it as finished software. The value is in learning: the code is open and free, so you can read it and see how a continuously working assistant is organised — when it should act, when it should stay quiet, and when it should ask a human.

If you want to try it, do so safely: use a throwaway repository, not your real project; use a separate GitHub and operating-system account for the assistant; and review every pull request it proposes before merging it. Agents that work on their own are a major industry direction — and understanding their rules now, while they are still experimental, is a skill that will matter as these tools enter everyday work.