AI Coding & Vibe Coding Published 2026-11-02

DeepSeek Super-Agent (dsh-superagent): What It Really Is, What It Needs, and What It Costs

How we checked: everything below comes from DeepSeek's own sources: the dsh-superagent-prototype and deepseek-harness repositories on GitHub (README, deployment and philosophy docs), the DeepSeek Harness page, and DeepSeek's API docs (Models & Pricing, Search & Fetch, Anthropic API guide), all read on November 2, 2026. We did not install or run it ourselves, so there are no speed or quality claims here. Prices are in US dollars. Sources are listed at the end. We have no affiliate relationship with DeepSeek.

The short answer

  • It's a research prototype, not a product. The repository's first line says it is a "research prototype for reference only, not an official DeepSeek product," and DeepSeek asks you to treat it as research code, not production-ready software. Some headlines have framed it as a Claude Code rival. That's not what DeepSeek itself says.
  • It isn't a terminal coding assistant you chat with. It's a proactive agent for one GitHub repository. It watches GitHub (and optionally Feishu, ByteDance's workplace chat app, called Lark outside China), picks up requests from people you allow, works on each one in its own session and folder, and answers where it was asked, for example with a pull request.
  • The software is free. Both the Super-Agent bundle and DeepSeek Harness, the agent app it runs inside, are MIT-licensed. What you pay for is the model you connect.
  • If you use DeepSeek's own API, deepseek-flash costs $0.30 per million input tokens (cache miss) and $1.20 per million output tokens at peak, and deepseek-v4-pro costs $1.32 and $3.96. Off-peak rates are half.
  • One new cost started November 1. Web Search and Fetch calls made through the DeepSeek API, including those made automatically by Harness, Claude Code or Codex, are now billed at $0.75 per 1,000 successful requests each.

What DeepSeek actually released

Two related things, and it helps to keep them apart:

DeepSeek Harness (dsh) dsh-superagent
What it is An open-source agent app with a Web UI and desktop app, built on an "everything is a plugin" design A plugin bundle for Harness: a proactive coding agent for one GitHub repo
Status Public preview (DeepSeek's site); the README calls it a developer preview with "compatibility-breaking changes" ahead Research prototype, "not an official DeepSeek product"
License MIT MIT
Price Free Free
How you start npx @deepseek-ai/dsh web (Web UI on 127.0.0.1:3080) A multi-step install on top of a dedicated Harness profile

DeepSeek also says the published prototype is not identical to its internal version, or to the design in its article Exploring Super-agent Design: Towards a More Autonomous Agent System. Some ideas in that article "are directions for future development and are not implemented in this prototype." So if you've read about the article's design, don't assume the code on GitHub does all of it.

The prototype repo was created on October 31, 2026 and had just over 12,000 GitHub stars when we checked on November 2.

How the Super-Agent works

The idea is an agent that lives alongside a team instead of inside one chat window:

  • A main session routes, workers do the work. One resident main session reads incoming messages and decides what becomes a work item (a "Ticket"). Each Ticket gets its own worker session and directory.
  • Sources. GitHub is the main source (new issues, new pull requests, comments, and default-branch CI results). Feishu is optional.
  • A local web panel on 127.0.0.1 shows only what needs a person.
  • It acts only when asked, or when a team-written rule allows it. A request counts when someone on your allowed list addresses the agent. Anything it would do on its own needs a rule, and it starts in a mode that only records what it would have done.

That last point is controlled by an accept mode:

Mode What happens
off Only explicit requests become work; nothing unrequested is even considered
observe (default) The agent assesses unrequested work and records "would accept," but accepts nothing
auto A positive assessment of unrequested work starts it automatically

DeepSeek's advice is to start in observe and switch to auto only after reviewing what it recorded. Explicit requests work in every mode.

What you need before you try it

From DeepSeek's deployment guide:

  • Node.js 22.19+ or 24+, pnpm, git 2.32+, Python 3.8+ and tmux.
  • GitHub CLI (gh) 2.48 or later.
  • A repository on github.com. GitHub Enterprise Server isn't supported. Install git-lfs if the repo uses Git LFS.
  • A separate GitHub user account for the agent, with write access to the repo. A GitHub App can't be the agent, because the prototype reads the account's notifications. The agent comments, pushes branches and opens pull requests as that account.
  • A model provider configured in DeepSeek Harness. The Super-Agent uses the profile's default model unless you set another.
  • Optional: lark-cli and a Feishu or Lark bot, if you want chat requests.

The install itself is a pinned global npm package (@deepseek-ai/[email protected]), a clone and build of the prototype repo, and a dedicated Harness profile called superagent. DeepSeek warns that the profile-creation command replaces an existing profile and deletes its machine configuration, so run it only once and never for upgrades. The README also suggests letting a coding agent walk you through SETUP.md, one question at a time.

The safety part you shouldn't skip

This is the section to read twice. DeepSeek's own deployment guide says both of the prototype's presets run with danger-full-access as your operating-system user. A worker "can execute programs by path and read the operator's home directory." The account wrappers only choose which credentials the agent uses. They don't isolate the process.

DeepSeek's recommendation is a dedicated OS account that contains only the agent's credentials. In practice, that means:

  • Don't run it under the same user account that holds your personal SSH keys, cloud credentials or browser profiles.
  • Give the agent's GitHub account access to the one repo it serves, not your whole organization.
  • Keep the accept mode on observe until you've read its recorded decisions.

What it costs to run

The software costs nothing. Your bill is model usage, plus web tools if the agent searches. If you connect DeepSeek's own API, these are the published rates per 1 million tokens:

deepseek-flash deepseek-v4-pro
Model version DeepSeek-V4.1-Flash-1017 DeepSeek-V4-Pro-0813
Input, cache hit (peak / off-peak) $0.006 / $0.003 $0.044 / $0.022
Input, cache miss (peak / off-peak) $0.30 / $0.15 $1.32 / $0.66
Output (peak / off-peak) $1.20 / $0.60 $3.96 / $1.98
Context / max output 1M / 384K tokens 1M / 384K tokens
Vision Yes No
Concurrency limit 2,500 500

When "peak" is. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. Everything else is off-peak, including weekends. In US Eastern Standard Time that's roughly 8–11 pm and 1–5 am, starting Sunday evening. For most North American daytime work, you're already on off-peak rates.

A rough example (our arithmetic, not a measurement). Say one agent task reads 2 million input tokens, 80% of them from cache, and writes 100,000 output tokens. At peak rates that's about $0.99 on deepseek-v4-pro and about $0.25 on deepseek-flash, or about half that off-peak. Real tasks vary widely, because an always-on agent rereads context often, and cache hits are what keep that cheap.

Search and Fetch. These are billed per successful request, $0.75 per 1,000 for each. Failed requests, empty search results and timeouts aren't billed, and the number of results you ask for doesn't change the price. DeepSeek notes that when you use its API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already hit these APIs. They were free until 00:00 Beijing time on November 1, 2026 (noon US Eastern on October 31). They're billed now.

How billing works. Usage is deducted from a topped-up balance or a granted balance, and the granted balance is used first. The pricing page doesn't list a free monthly API allowance, and DeepSeek reserves the right to change prices.

Super-Agent vs Claude Code and Codex

They solve different problems, so this isn't a like-for-like price fight:

  • Claude Code and Codex are mainly interactive coding agents that come with a chat subscription (full access starts at $20 a month), used in a terminal, IDE or app. We compare them in Claude Code vs Codex.
  • dsh-superagent is a team-facing, always-on agent that sits on a GitHub repo and responds to issues, PRs and CI. There's no subscription, just pay-per-token model costs, and it's research code you host and secure yourself.

There's also an overlap worth knowing about. DeepSeek's API speaks the Anthropic format at https://api.deepseek.com/anthropic, and DeepSeek documents using it inside Claude Code. When you do that, Claude model names are mapped: claude-opus* goes to deepseek-v4-pro (billed at the V4 Pro price), and claude-sonnet* and claude-haiku* go to deepseek-flash. That's a way to get DeepSeek pricing in a familiar tool without running the Super-Agent at all. The Search and Fetch charges above still apply there.

Who should try it

  • Try it if you maintain a GitHub repo, you're comfortable with Node, tmux and the GitHub CLI, and you want to experiment with an agent that triages issues and opens PRs. Plan to set up a separate OS user and a separate GitHub account for it.
  • Skip it for now if you want a polished coding assistant to use today, you need GitHub Enterprise Server, or you can't isolate an agent that runs with full access to its user account. A subscription tool like GitHub Copilot, Claude Code or Codex is the easier path.

FAQ

Is DeepSeek Super-Agent an official DeepSeek product? No. DeepSeek's repository calls dsh-superagent a "research prototype for reference only, not an official DeepSeek product."

Is DeepSeek Super-Agent free? The code is free under the MIT license, and so is DeepSeek Harness. You pay for whatever model provider you configure. With DeepSeek's API, that's per-token pricing plus $0.75 per 1,000 successful Search or Fetch requests.

Does it work with GitLab or GitHub Enterprise Server? DeepSeek's guide requires a repository on github.com and says GitHub Enterprise Server isn't supported. Its built-in sources are GitHub and, optionally, Feishu.

Will it make changes nobody asked for? Only if you allow it. Explicit requests come from people on your allowed list. Unrequested work needs a team-written rule, and the default observe mode only records what it would have done. Switching to auto lets it start that work on its own.

Why did my DeepSeek bill go up in November when I use Claude Code? Since 00:00 Beijing time on November 1, 2026, server-side Search and Fetch calls made through the DeepSeek API in Harness, Claude Code or Codex are billed at $0.75 per 1,000 successful requests each. Before that they were free.

Sources

Tools in this article

Get new AI tools in your inbox

A short digest of verified new tools and price changes. No spam.