DeepSeek Super-Agent: Free Open-Source Coding Agent

Updated: November 2026 · 5 min read · By the CreatorSprout team

Launch coverage based on DeepSeek's October 31, 2026 open-source release and reporting by Bloomberg, Startup Fortune and independent reviewers — we have not tested this prototype hands-on. It is research code, not a finished product. No affiliate links in this article.

Most coding agents today are polite assistants: you open a chat, you describe the bug, you watch it work. Super-Agent flips that relationship. The idea, laid out in DeepSeek's accompanying design writeup Exploring Super-agent Design: Towards a More Autonomous Agent System, is an agent that stays on duty across conversations and tools, keeps context across tasks, and acts when the rules allow it — even when nobody tagged it. If that sounds powerful, it is. If it sounds like something you should not point at production code yet, that is also correct, and this article keeps both halves in view.

What was actually released

DeepSeek's Harness team published the repository dsh-superagent-prototype under the DeepSeek GitHub organisation on October 31, 2026, with a single commit adding the prototype. It is a plugin bundle for DeepSeek Harness, the open-source, self-hostable agent framework DeepSeek released on August 13, 2026, also MIT-licensed. Reports put the prototype at roughly 2,000 stars about 12 hours after release and roughly 4,100 stars by the morning of November 1 — treat those as reported figures, not a live count, but the direction is clear: developers want open coding-agent infrastructure they can run themselves.

What the prototype does, in plain terms: it is scoped to one GitHub repository. It listens to GitHub — and optionally to Feishu chat — and when an authorised person asks, or a team-written rule allows it, it opens a work item, works in its own agent session and directory, and answers back in the original conversation, for example with a pull request. DeepSeek's own framing for the internal version goes further — tracking an issue from an instant message into GitHub, and monitoring public developer discussions for new information — but the published code is an adaptation, not that internal system.

How it works: Router and Workers

The design splits the agent in two, and the split is the whole idea:

The Router holds the broader context and handles communication. It decides which permitted requests become work, and routes them onward.
Workers execute separate coding tasks in parallel, each in its own session and directory, then return results — a fix, a pull request — to the original conversation.

The writeup adds two more ambitions around that core: speculative execution (starting work before it is certain to be needed) and 24/7 persistence (an agent that stays on duty rather than waking per chat). Be careful with both: reviewers note that some designs in the paper are future directions, not implemented in the published prototype. Judge the release by the repository, not by the paper's ambitions — and judge the paper as a direction of travel, not a feature list.

A concrete example from DeepSeek's description shows why this matters to a team: someone reports a bug in a group chat, the agent identifies the issue, attempts to reproduce it, and submits a pull request if it finds a solution — without anyone explicitly summoning it. That is the difference between an assistant and a colleague, and it is also why the permission rules matter more here than in a chat-based tool.

What it costs

There is no licence fee and no subscription: the prototype is MIT-licensed, and Harness is self-hostable, so you can download, run and inspect the code for free. That is the complete pricing story for the software itself — and it is not the complete cost story for running it.

Model usage: you must configure a model provider, and that provider bills its own API usage. DeepSeek attaches no price, SKU or subscription to Super-Agent itself, so any running cost depends on the provider you choose — we do not quote a figure here because there is none to quote.
Your time: this is preview software on an alpha framework. Expect setup friction, uneven code style (DeepSeek's README warns of it) and bugs, and budget review time for every change it proposes.

Compare that with the closed alternative: tools like GitHub Copilot Fusion bill per model used, while Super-Agent bills nothing for the agent — but hands you the hosting, the model bill and the responsibility. Free code is not free operation.

What you need to run it

The requirements, as documented for the prototype, are specific — this is a developer setup, not a download-and-double-click tool:

A dedicated DeepSeek Harness profile, on Harness 0.2.1-alpha.2
A GitHub account reserved for the agent (not your personal account)
A configured model provider
Node.js 22.19 or later, pnpm, git, tmux, Python 3.8 and the GitHub CLI

The separate GitHub account is not optional colour — it is the permission boundary. The agent acts as that account, so what it can touch is defined by what that account can touch. Keep it scoped to one throwaway repository first.

Super-Agent vs Claude Code

The comparison DeepSeek itself invites is with Anthropic's Claude Code, and the timing sharpens it: the release landed weeks ahead of Anthropic's expected IPO roadshow. The honest version of the comparison is narrower than the headlines:

Initiative: Claude Code waits for instructions in a conversation. Super-Agent is built to act on rules and observations without a prompt.
Openness: Super-Agent's rulebook can be read and rebuilt — it is MIT code you can inspect. Claude Code is a supported commercial product with a support desk.
Maturity: this is where the comparison flips. Claude Code is a product; Super-Agent is expressly not one. The only one of the two DeepSeek declares not ready for production is its own.

For a solo creator or small team, the practical read is: Claude Code (and the broader agent wave covered in DeepSeek's own V4.1 Flash release) is what you ship work with today; Super-Agent is what you study to see where autonomous coding agents are heading, and experiment with on code you can afford to break.

The honest caveats

We haven't tested it. Everything here comes from the repository description, DeepSeek's writeup and independent coverage of the October 31 release. No hands-on claims are made.
Research code, by the author's own label. The README's first line frames it as a research prototype for reference only, not an official DeepSeek product — and the code differs from both the internal version and the paper's implementation.
Workers run commands without approval, as your OS user. Reviewers flag this directly: with research code, you vet the code, configure the sandbox and review every change. There is no support desk, and a sandbox-escape flaw in the preview Harness was only recently fixed, per independent coverage.
Benchmarks are DeepSeek's own simulation. The design writeup's simulation figures are not independent benchmarks, and no consumer product, API SKU or pricing is attached to this release.
Star counts are a signal, not a verdict. Roughly 4,100 stars in a day shows developer interest in open agent infrastructure; it says nothing about whether this prototype is safe or correct on your repository.

Verdict

DeepSeek Super-Agent is worth your attention as architecture, not yet as a tool you trust with real code. A free, open, proactive coding agent from a frontier lab — one that watches GitHub, works across conversations and answers with pull requests — is a genuinely major direction for developer tools, and the open code means anyone can inspect, fork and learn from it. But DeepSeek is unusually honest about the gap between the idea and this release: research prototype, alpha framework, code that differs from the paper. Read it, run it on a throwaway repo, steal the Router-and-Worker idea for your own workflows — and keep it far away from production until a supported version exists.

FAQ

Is DeepSeek Super-Agent free?

The prototype code is free under the MIT licence, and DeepSeek Harness is self-hostable, so there is no licence or subscription fee. You still need a configured model provider, and that provider's API usage is billed under its own terms.

Is Super-Agent a finished DeepSeek product?

No. DeepSeek's own README calls it a research prototype for reference only, not an official DeepSeek product, and not production-ready. The published code also differs from the company's internal version and from its design paper.

How is Super-Agent different from Claude Code?

Claude Code waits for you to ask. Super-Agent is designed to be proactive: it monitors GitHub (and optionally Feishu chat), spots a coding problem, and works toward a fix — for example, opening a pull request — without being explicitly summoned, within rules the team sets.

What do you need to run Super-Agent?

A dedicated DeepSeek Harness profile, a separate GitHub account reserved for the agent, a configured model provider, and Node.js 22.19 or later, pnpm, git, tmux, Python 3.8 and the GitHub CLI. It runs on DeepSeek Harness 0.2.1-alpha.2, which is preview software.

Should you let Super-Agent loose on a real repository?

Not yet, and not unsupervised. The prototype is research code, workers run commands as the operator's OS user, and reviewers flag the preview Harness as unaudited. Try it on a throwaway repository first, review every change, and keep it away from production code and secrets.