DeepSeek Super-Agent Is Free Code, Not a Finished Product
On October 31, DeepSeek's Harness team published a prototype that tries to flip the usual coding-agent workflow: instead of waiting for a developer to prompt it, the agent watches a GitHub repository, picks up permitted work, and answers in the original conversation, for example with a pull request. The code is free under the MIT licence. The important qualifier is in DeepSeek's own wording: it is a research prototype for reference only, not an official DeepSeek product, and it is not production-ready software.
What the prototype actually does
The repository is deepseek-ai/dsh-superagent-prototype, packaged as a bundle for DeepSeek Harness. It is scoped to one GitHub repository at a time. It listens to GitHub and, optionally, Feishu, the workplace chat app sold as Lark outside China. A Router evaluates permitted requests arriving from those channels, while isolated Worker sessions carry out accepted tasks and return results to the conversation where the work began.
That design is the whole point. DeepSeek's accompanying essay, published on its official WeChat account on the evening of October 31, Beijing time, frames the goal as using as little human attention as possible while delivering as much value as possible. In DeepSeek's internal use, the team says its own Super-agent reproduces and fixes bugs, responds to code reviews, and helps build DeepSeek Harness itself. The public prototype is narrower, and DeepSeek says some designs in the essay are directions for future development, not features implemented in the published code.
What it does not do
First, this is not a no-sign-up chatbot and not a consumer product you open in a browser. It is code for developers to inspect, adapt, and run. Second, free code is not free running cost: the code itself is free under the MIT licence, but you pay for whatever model access you connect. Third, do not treat the public prototype as DeepSeek's internal agent. DeepSeek says the published code differs from its internal system, that much of the engineering was done by coding agents, and that the research code may still contain bugs.
The performance numbers need the same caution. In DeepSeek's own simulated tests, reported by secondary coverage, half of roughly 20 work requests finished within 12.4 minutes and the slowest 5% of tasks fell to 15.7 minutes from 29.5 minutes. Those are DeepSeek's own simulation figures, not an independent benchmark of your repository, your model, or your workflow.
Who should try it, and how
This is for engineers and teams who want to study or experiment with proactive agents, not for anyone looking for a finished assistant. If that is you, start small:
1. Read the README first. Check what the prototype listens to, what counts as a permitted request, and what it is allowed to change before you point it at anything you care about.
2. Use a throwaway repository. Let it propose work there, review every pull request, and see how it behaves when a request is vague, out of scope, or wrong.
3. Keep a human in the loop. Budgets, approvals, and review are not optional extras for an agent that acts without being tagged. The prototype's value is in the proposals it surfaces; the decision to merge is still yours.
If you want a coding agent you can use today rather than a research prototype to study, start with our tested free coding agents instead. For the broader directory, see our agents directory, and for the API side of the same workflow, see our free APIs page.
FAQ
Is DeepSeek Super-agent free?
The published code is free under the MIT licence. Model access, hosting, and any other services you connect to it are separate costs.
Does it work without an account?
Not in the way a no-sign-up chatbot does. It is self-hosted developer code that watches a GitHub repository, not a web chat product you open without logging in.
Can I run it on a production repository today?
DeepSeek advises against treating it as production-ready. Try it on a throwaway repository first, review everything it proposes, and adapt the code before wider use.
Where is the source?
The repository is deepseek-ai/dsh-superagent-prototype. Secondary coverage that helped verify the October 31 release includes MadRobot's November 1 report and a DEV roundup of tools launched October 29–31.