AgentComparison
DeepSeek (DeepSeek Harness Team) · product · coding agent prototype (self-hosted, DeepSeek Harness bundle)

DeepSeek dsh-superagent

DeepSeek's open-source prototype of a proactive coding agent for one GitHub repository, built as a DeepSeek Harness bundle.

Group: Coding agents

dsh-superagent is a proactive coding agent for one GitHub repository, published 31 October 2026 as an open-source bundle for DeepSeek Harness (DSH). It listens to GitHub and, optionally, Feishu; when someone on an allow-list asks for work, it opens a work item, works in its own agent session and directory, and answers in the original conversation, for example with a pull request. A web panel shows what needs a person. The vendor is explicit that this is a research prototype for reference only, not an official DeepSeek product and not production-ready software.

Documentation researched, not testedChecked 1 November 2026Vendor documentation, not a test

The decision in plain English

The code is free under MIT, but two qualifiers matter more than price. First, the vendor's own README calls it research code, not an official DeepSeek product, and says to review and adapt it before use; press coverage describing it as a launched DeepSeek product overstates what the repository claims. Second, it runs on DeepSeek Harness with a model provider you configure yourself, so running cost depends entirely on your provider and workload and is not stated in the repo. Setup is deliberately agent-led (a coding agent reads SETUP.md and interviews you), needs Node.js 22.19+ or 24+, pnpm, git 2.32+, tmux, Python 3.8+, GitHub CLI 2.48+ and DSH, and defaults are conservative: a dedicated GitHub agent account, an allow-list of logins who may ask for work, and people merging its pull requests until they have been right for a while.

Best fit

Teams already running DeepSeek Harness who want to study or adapt a working example of a proactive, low-attention coding agent for a single repository, and are comfortable operating research-grade self-hosted code.

Poor fit

Anyone who wants a supported, production-ready coding agent or a managed service: the vendor states this is research code with possible bugs and inconsistent engineering practices, and there is no hosted version, support channel or published release.

Documented capabilities

Vendor statements, not verified task outcomes.

  • Proactive coding agent for one GitHub repository: opens a work item when an allowed person asks, works in its own session and directory, and answers in the original conversation, for example with a pull request Source · checked 2026-11-01
  • Open source under the MIT licence Source · checked 2026-11-01
  • Optional Feishu listener alongside GitHub Source · checked 2026-11-01
  • Web panel showing what needs a person Source · checked 2026-11-01

Permissions and human control

  • Acts only when a person on a configured allow-list asks, or when a rule the team wrote allows it; the design principle is to take as little human attention as possible.
  • The vendor recommends a dedicated GitHub account for the agent, never the operator's own: everything it writes appears under that name, and its commits name the agent account as author.
  • Auto-merge on the agent's own pull requests is off by default; the agent never enables it as the agent account, so people merge until its PRs have proven reliable.

Documented controls do not prove that an agent always follows them.

Limits and open questions

  • Vendor disclaimer: research prototype for reference only, not an official DeepSeek product and not production-ready; review and adapt before use.
  • Runs only as a bundle for DeepSeek Harness, which itself needs a configured model provider; model cost is not stated in the repo.
  • One GitHub repository per setup; Feishu support is optional and needs lark-cli.
  • No releases published; installed from source with Node.js, pnpm, tmux, Python 3 and GitHub CLI prerequisites.
  • Described by the vendor; not tested by AgentComparison

Privacy and security

Self-hosted on your own machine or server and works against GitHub accounts and repositories you control; the agent acts under its own dedicated GitHub account. The repository does not publish a data-handling, telemetry or retention policy, and the vendor advises reviewing and adapting the code before use. Vendor documentation, not independently audited.

Feature-by-feature: what is documented and what is not

Official sources checked 1 November 2026. A documented feature is a vendor statement, not proof of success.

FeatureEvidence statusDetail and limit

How to start with DeepSeek dsh-superagent narrowly

  1. Ask your coding agent to read SETUP.md in the repository and set up dsh-superagent with you; it asks one short question at a time and shows the full plan before executing.
  2. Have the prerequisites ready: Node.js 22.19+ or 24+, pnpm, git 2.32+, tmux, Python 3.8+, GitHub CLI 2.48+, and DeepSeek Harness with a model provider configured.
  3. Create a dedicated GitHub account for the agent; never let it act as your own account.
  4. Keep the allow-list of logins who may ask for work short at first, and let people merge the agent's pull requests until they have been right for a while.

Official product/setup source. These are documentation-based steps, not a hands-on setup test.

Common questions about DeepSeek dsh-superagent

What is DeepSeek dsh-superagent for?

Teams already running DeepSeek Harness who want to study or adapt a working example of a proactive, low-attention coding agent for a single repository, and are comfortable operating research-grade self-hosted code. The vendor describes it as: dsh-superagent is a proactive coding agent for one GitHub repository, published 31 October 2026 as an open-source bundle for DeepSeek Harness (DSH). It listens to GitHub and, optionally, Feishu; when someone on an allow-list asks for work, it opens a work item, works in its own agent session and directory, and answers in the original conversation, for example with a pull request. A web panel shows what needs a person. The vendor is explicit that this is a research prototype for reference only, not an official DeepSeek product and not production-ready software.

Who should look elsewhere?

Anyone who wants a supported, production-ready coding agent or a managed service: the vendor states this is research code with possible bugs and inconsistent engineering practices, and there is no hosted version, support channel or published release.

How much does DeepSeek dsh-superagent cost?

The dsh-superagent repository (read 1 November 2026) is MIT-licensed with no releases and no paid tier. It runs on DeepSeek Harness, which needs a model provider; the repo does not state which providers or what usage costs apply.

What can it access and what needs my approval?

Acts only when a person on a configured allow-list asks, or when a rule the team wrote allows it; the design principle is to take as little human attention as possible. The vendor recommends a dedicated GitHub account for the agent, never the operator's own: everything it writes appears under that name, and its commits name the agent account as author. Auto-merge on the agent's own pull requests is off by default; the agent never enables it as the agent account, so people merge until its PRs have proven reliable.

What happens to my data?

Self-hosted on your own machine or server and works against GitHub accounts and repositories you control; the agent acts under its own dedicated GitHub account. The repository does not publish a data-handling, telemetry or retention policy, and the vendor advises reviewing and adapting the code before use. Vendor documentation, not independently audited.

Has AgentComparison tested it?

No. This record is built from the vendor's own documentation, checked 2026-11-01. Vendor claims are attributed, not confirmed.

How to evaluate DeepSeek dsh-superagent yourself

  1. Pick one low-risk task that matches its documented capabilities: proactive coding agent for one github repository: opens a work item when an allowed person asks, works in its own session and directory, and answers in the original conversation, for example with a pull request.
  2. Write the expected output and what counts as failure before you start.
  3. Connect the minimum access needed, and check approval settings first.
  4. Run it twice and compare, then review every action it took.
  5. Check cost after any free allowance. Unknown price is not free.

Related

Documented use cases: Software development and repository work.

Sources and history

Change history

Documentation overview expanded with current pricing/access sources. No hands-on results claimed.

Report a correction