DeepSeek Super-Agent vs Claude Code: What's Actually Shipping
DeepSeek's Super-Agent is a free, open-source research prototype, not a finished product. Here is how it actually compares with Anthropic's commercial Claude Code.

DeepSeek's newly open-sourced "Super-Agent" and Anthropic's Claude Code are not the same kind of thing. One is a free, MIT-licensed research prototype you self-host and wire up yourself. The other is a commercial, subscription- or API-metered product with official installers, an IDE presence, and a support organization behind it. Below is what each actually is, what each is documented to do, and who should realistically reach for which one right now.
- DeepSeek Super-Agent: open-source research prototype, repo
deepseek-ai/dsh-superagent-prototype, published October 31, 2026, MIT license, free to self-host (model API costs apply separately). - Claude Code: Anthropic's commercial agentic coding tool, available via Claude Pro ($20/mo), Max ($100/mo), Team, Enterprise, or Anthropic API usage.
- DeepSeek's own repository calls Super-Agent "a research prototype... for reference only, not an official DeepSeek product."
- No independent, head-to-head benchmark comparing the two currently exists — this piece compares documented features and access models, not scores.
What DeepSeek's Super-Agent actually is
DeepSeek Super-Agent, formally dsh-superagent, is a single GitHub repository that DeepSeek's "Harness Team" pushed to the public deepseek-ai organization on October 31, 2026, under an MIT license. The repository's own description frames it plainly: it is "a proactive coding agent for one GitHub repository, packaged as a bundle for DeepSeek Harness." A disclaimer at the top of the README goes further, describing the code as a research prototype released "for reference only" and explicitly stating it is "not an official DeepSeek product."
In practice, it is a bundle that plugs into DeepSeek Harness, DeepSeek's broader open-source agent runtime released earlier in 2026. Super-Agent itself doesn't ship a model — the README does not name which model it runs on, only that it is packaged for Harness. Anyone standing it up supplies their own API access and credentials.
What Claude Code actually is
Claude Code is Anthropic's own description, from its official documentation: "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools." It ships as a native terminal CLI, a VS Code and JetBrains extension, a desktop app, and a browser-based version at claude.ai/code, all backed by the same underlying engine so that a project's CLAUDE.md instructions, settings, and MCP server connections carry across surfaces. It has been a maintained, versioned Anthropic product since 2025, with its own documentation site, release channel, and enterprise support tier — a different category of thing than a one-commit research drop.
Access model and pricing
This is where the two diverge most sharply. Super-Agent's code costs nothing: MIT licensing means anyone can clone, modify, and run it without a software fee. But it is not free to operate, because it still needs a model behind it through DeepSeek Harness, and that model access is billed separately through whichever API the operator connects.
Claude Code is bundled into Anthropic's subscription tiers rather than sold on its own. Per Anthropic's current pricing page: the Pro plan is $20/month billed monthly (or $17/month billed annually) with baseline usage limits; Max is $100/month with five times the Pro usage allowance; Team plans run $25/seat/month (standard) or $125/seat/month (premium) billed monthly, covering 2–150 users; and Enterprise starts at $20/seat/month with a 20-seat minimum plus usage billed at API rates. The free Claude plan does not include Claude Code at all — some paid tier or an Anthropic Console/API account is required.
| Dimension | DeepSeek Super-Agent | Claude Code |
|---|---|---|
| Status | Open-source research prototype, explicitly "not an official DeepSeek product" | Commercial, officially supported Anthropic product |
| License / cost of code | MIT, free to clone and modify | Proprietary; no public source release |
| How you pay | Free software, but you supply and pay for your own model API access via DeepSeek Harness | Claude Pro, Max, Team, or Enterprise subscription, or Anthropic API usage billing |
| Entry price | $0 for the code itself | $20/mo (Pro, monthly) or $17/mo (Pro, billed annually) |
| Install / setup | Self-hosted; README directs you to have a coding agent follow SETUP.md, or configure manually via docs/006-deployment.md | One-line installer script, Homebrew, WinGet, or Linux package managers; also VS Code, JetBrains, desktop app, and browser |
| Scope as shipped | Works one GitHub repository per deployment, triggered by issues/requests from allowed people | Works across local or remote repos, terminal, IDE, CI, Slack, and the web, with the same engine everywhere |
| Parallel/multi-agent work | Not documented in the public README beyond isolated per-request work sessions | Documented subagents (up to 20 concurrent by default, 3 levels of nesting), background agents, and agent teams coordinated by a lead agent |
| Named model backing | Not stated in the public repository | Anthropic's Claude model family, selectable per subagent or session |
| Support | Community/GitHub issues only; authors call it "research code," "not production-ready," possibly containing bugs | Official docs, Claude Academy courses, enterprise support tiers |
Underlying models
Claude Code is explicit about what powers it: Claude models, with the ability to assign a specific model (for example, naming "sonnet" in a subagent's configuration) per task or subagent. Super-Agent's repository, by contrast, does not name a specific model. It is "packaged for DeepSeek Harness," meaning the model is whatever the operator connects through Harness — the public prototype is a harness/orchestration layer, not a model release in itself.
Setup and installation
Claude Code installs with a single command — curl -fsSL https://claude.ai/install.sh | bash on macOS/Linux/WSL, an equivalent PowerShell one-liner on Windows, or a Homebrew/WinGet/apt/dnf package — followed by claude in any project directory and a login prompt. Native installs auto-update in the background.
Super-Agent's setup is aimed at a technical operator willing to configure infrastructure. The README's recommended path is to ask a coding agent to read the repo's own SETUP.md and walk through setup interactively; a manual path is documented separately in docs/006-deployment.md. It needs its own GitHub identity with repository write access and, per the project's numbered docs (covering domain, authorization, execution, prompts, the web panel, deployment, and architecture), a fair amount of configuration before it does anything useful. There's no equivalent of a one-line consumer installer.
That gap matters for who can realistically evaluate each tool this week. Claude Code's installers are built for someone who has never configured a background service: download, authenticate, run claude in a project folder, done. Super-Agent instead assumes an operator who already understands GitHub app permissions, process management, and how to wire a model API behind DeepSeek Harness — closer to standing up a self-hosted CI runner than installing a desktop tool. Neither approach is wrong for its intended audience, but they are not interchangeable starting points.
Autonomy features, as documented by each maker
Super-Agent's stated design goal, quoted from its own README, is to "take as little human attention as possible, and give as much value as possible." Mechanically, it listens for requests — on GitHub, and optionally via the Feishu chat app — from people a project has explicitly authorized, opens an isolated agent session and working directory per request, and replies in the original thread, typically with a pull request. A web panel surfaces anything that still needs a human decision. The agent is scoped to act "only when asked or when a team rule allows it," per the README.
Claude Code's documented autonomy features are broader in scope: it can plan and execute multi-file changes, create commits and branches and open pull requests directly, connect to external tools via MCP, run on a recurring schedule through cloud or desktop scheduled tasks, be piped and scripted from the command line for CI use, and — notably for direct comparison with Super-Agent's single-repo, single-trigger design — spawn multiple subagents that a lead agent coordinates and merges, plus fully separate background agents and project-level "agent teams" for larger, multi-session bodies of work.
Anthropic's documentation also spells out the mechanics of that coordination in more detail than DeepSeek's prototype currently does for its own isolated sessions: a lead agent writes a delegation message summarizing each subtask, independent subagents can research or edit in parallel, and results flow back to the main conversation for Claude to synthesize rather than requiring a separate manual merge step. Concurrency and nesting are both bounded and configurable (20 concurrent subagents and three levels of nesting by default, per Anthropic's docs) rather than left undocumented.
Documented limitations
DeepSeek's own README is unusually candid about Super-Agent's rough edges: it calls the code "research code," says it is "not production-ready," warns it "may contain bugs," notes that coding style and engineering practices "may be inconsistent," and states plainly that the published version differs from DeepSeek's internal version — with some of the designs described in the project's own documentation not actually implemented in what was released. It explicitly asks anyone using it to review and adapt the code for their own environment first.
Claude Code's public documentation does not carry an equivalent blanket disclaimer, consistent with its status as a maintained commercial product, though Anthropic's docs do note usage limits tied to each subscription tier (for example, the default 20-concurrent-subagent ceiling, which is configurable) and that most surfaces require a paid Claude subscription or Anthropic Console account to use at all.
Claims still unconfirmed by DeepSeek itself
Coverage of the Super-Agent release has included claims that go beyond what the repository states. Some reports have described the agent as able to monitor developer group chats on its own and surface or act on problems without being directly asked, and some secondary coverage has implied a hosted, API-accessible version might follow. Neither claim appears in DeepSeek's own repository or README: the documented behavior is limited to responding to requests from explicitly authorized people via GitHub or an optionally connected Feishu channel, and the project ships as self-hosted code with no hosted web or API product mentioned anywhere in the repo. Readers should treat the broader "watches your team chat" and "hosted access coming" framing as reported by secondary outlets and unconfirmed by DeepSeek, not as DeepSeek's own stated claim.
Separately, some outlets have published informal, single-run speed or output comparisons between the two tools. None of this amounts to an independent, reproducible benchmark, and no such benchmark was found for this article. Any claim that one tool is categorically "faster" or "better" at coding tasks than the other, absent a disclosed, repeatable methodology, should be read as anecdote rather than evidence.
Who each is realistically for right now
Super-Agent, as published, suits a hobbyist, researcher, or platform engineer who wants to study or extend an open proactive-agent design, already runs DeepSeek Harness or is willing to, and is comfortable debugging research-grade code without vendor support. It is not, by its own authors' description, something to point at a production repository today without review.
Claude Code suits an individual developer or team that wants a supported product today: a one-line install, IDE and CI integrations, documented concurrency and nesting limits for multi-agent work, and a predictable subscription or metered-API bill. For teams already comparing agentic coding tools more broadly, Pandromeda's side-by-side look at Claude Code against OpenAI's Codex and its comparison with Google's Gemini CLI cover how it stacks up against other shipped commercial agents, while a primer on what Claude Code is and how it works is a useful starting point for readers new to the product.
What's next
Super-Agent is a prototype, not a roadmap commitment, so there's no confirmed timeline for it becoming an official DeepSeek product, gaining a named model, or shipping any hosted version — anyone interested should watch the deepseek-ai/dsh-superagent-prototype repository directly rather than secondary coverage for actual changes. Claude Code, as an actively maintained product, continues to add surfaces and orchestration features (agent teams and dynamic workflows among the most recent), documented through Anthropic's own changelog and docs site. Readers evaluating either should treat this as a snapshot of two very differently staged projects on the date of publication, not a permanent state of the market.
Frequently asked questions
What is DeepSeek Super-Agent exactly?
It is an open-source research prototype in the public GitHub repository deepseek-ai/dsh-superagent-prototype, published October 31, 2026 under an MIT license. DeepSeek's own README calls it a research prototype released 'for reference only' and states it is 'not an official DeepSeek product.'
Is DeepSeek Super-Agent free to use?
The code itself is free under the MIT license. Running it still requires a model behind it through DeepSeek Harness, and that model access is billed separately through whichever API the operator connects, so there is an ongoing cost even though the software has none.
Does DeepSeek Super-Agent monitor group chats on its own?
DeepSeek's own repository describes it as responding to requests from explicitly authorized people via GitHub, with optional Feishu chat support. Broader claims that it autonomously watches team chats for problems and acts without being asked have appeared in secondary reporting but are not confirmed by DeepSeek's own documentation.
What models power Claude Code?
Claude Code runs on Anthropic's Claude model family. A specific model can be assigned per task or per subagent in its configuration, according to Anthropic's official documentation.
How much does Claude Code cost?
Claude Code is included with a Claude Pro subscription ($20/month billed monthly, or $17/month billed annually), Max ($100/month), Team ($25 or $125 per seat/month), or Enterprise plans, or through Anthropic API usage billing. It is not available on Claude's free plan.
Is there a benchmark comparing DeepSeek Super-Agent and Claude Code?
No independent, reproducible benchmark comparing the two was found. Some outlets have published informal single-run comparisons, but these are anecdotal rather than rigorous benchmarks, and should be treated accordingly.
Sources
- DeepSeek: dsh-superagent-prototype README (GitHub)github.com
- DeepSeek: dsh-superagent-prototype LICENSE (GitHub)github.com
- Anthropic: Claude Code overview documentationcode.claude.com
- Anthropic: Claude Code subagents documentationcode.claude.com
- Anthropic: Claude pricingclaude.com
- Anthropic: Claude Code product pageclaude.com
Felix Moreau writes Pandromeda's software coverage and how-to guides. He covers Windows, macOS and Linux updates, the apps people rely on, emulators and developer tools, and turns official documentation into clear, numbered steps that work on the current version.


