DeepSeek open-sources Agent framework, tens of thousands of Stars in one day, directly targeting Claude Code
On August 13, DeepSeek did two things: the official release of the V4 Pro model, and open-sourcing its Agent framework Harness (abbreviated as dsh) under the MIT license. On the day the repository went public, its Star count surged to 2.
💡 What You Will Learn
On August 13, DeepSeek did two things: the official release of the V4 Pro model, and open-sourcing its Agent framework Harness (abbreviated as dsh) under the MIT license. On the day the repository wen
📜 Table of Contents
DeepSeek Open-Sources Agent Framework, Tens of Thousands of Stars in a Day, Directly Taking on Claude Code
On August 13, DeepSeek did two things: released the official version of the V4 Pro model, and open-sourced its Agent framework Harness (dsh for short) under the MIT license.
On the day the repo went public, star count hit 28,000, and by the time of writing, it had surpassed 70,000. Forks exceeded 6,000. Jiqizhixin, QbitAI, and 36Kr all covered it that same day.
I use Agent frameworks daily for work, so I pulled the code and ran it the moment I saw the news. Here's my take.
What Exactly Is dsh
dsh is an Agent Harness — the "body" for AI coding agents: how to call models, execute commands, read/write files, handle permission approvals, and do sandbox isolation.
The benchmark is clear: Anthropic's Claude Code, OpenAI's Codex CLI.
DeepSeek previously only had models, no Agent entry point of its own. If users wanted to write code with DeepSeek, they had to plug into third-party tools like OpenCode or Kilo Code, hoping someone else's framework would play nice with their model. dsh fills that gap.
Getting started is as simple as one command:
npx @deepseek-ai/dsh web
Open 127.0.0.1:3080 in your browser — model config, workspace selection, sessions, approvals, and plugin management are all clickable in the web UI. This sets it apart from both Claude Code and Codex, which prioritize the terminal. dsh goes straight for a local Web UI.
133 Plugins, Only 9 Core Files
dsh's biggest difference lies in its architecture. The official slogan is "Everything is a Plugin," running on a plugin microkernel called Cordis.
A developer on Juejin dumped a local instance and found 133 plugins assembled in total, each individually startable/stoppable: model adapters, tool registration, session logging, sandbox policies, permission approvals, Claude Code hooks bridging, MCP, ACP, web search — all plugins. The core source code is just 9 files, and even the Agent Loop itself is a plugin that can be swapped out entirely.
This is the complete opposite of Claude Code's approach. Claude Code uses a Rust monolith core, with MCP tools, hooks, and skills all hanging off the outside — you can add things, but you can't modify the backbone. Codex is similar, with model and product tightly coupled.
QbitAI's analogy hits the mark: dsh takes the "Android" route — ultra-thin kernel, everything replaceable, community can modify freely; Claude Code and Codex take the "iPhone" route — closed but with a polished, integrated experience.
Both approaches have trade-offs. A Rust monolith core has advantages in speed and controllability; a plugin framework adds a layer of abstraction. But dsh's openness means: don't like the built-in sandbox? Swap in a different plugin. Want to connect any OpenAI-compatible model endpoint? Add one line in settings. Want to call Claude Code itself as a sub-agent? dsh supports that natively.
Model-Agnostic, Four Modes to Switch Between
dsh doesn't lock you into DeepSeek's own models by default. In settings, DeepSeek official is just one provider — any OpenAI-compatible endpoint can be added as a custom provider, taking effect on save without a restart.
Four built-in working modes: Standard, PTC (Programmatic Tool Calls), Minimal, and Creator. Switching modes is like swapping plugin manifests — tools, prompts, and capabilities all change. That's a fundamentally different approach from tweaking a few config options in Claude Code.
The official release ships with 100+ built-in plugins, and the community has already produced third-party plugins like task boards, Git graphs, whale-girl skins, and stock ticker marquees. An awesome-deepseek-harness list has also been set up.
Plenty of Controversy Too
Let me be clear: alongside the praise, there's noise.
During the beta phase, reports surfaced of requests for ID card copies, and well-known cloud-native developer gaocegege publicly criticized privacy compliance. On HN, people questioned "plugin fatigue": do we really need to learn another new ecosystem? Who's responsible for npm supply chain security? The sharpest comment: "Everything is a plugin — did they just discover Unix pipes?"
The official team is honest about it too — the README prominently states "Developer preview, breaking changes ahead." It's still a ways off from "install and everything just works."
Lead developer Cui Tianyi spent nine years as a quant at Jane Street, confirmed team formation in May, and was still posting job listings in June saying "severely understaffed, interviewing every day." Two months later, this is what they delivered — execution is genuinely impressive.
My Take
dsh is worth watching, but it's too early to replace your daily driver with it today. A developer preview means APIs can change at any time, the plugin ecosystem is just getting started, and supply chain security plus permission models need time to prove themselves.
What's really worth tracking is the direction it represents: model vendors building Agents are starting to choose a third path. Previously, it was either programming assistants like Claude Code or "AI coworker"-style products. DeepSeek chose to build the foundation, handing assembly power to the community, and using model neutrality to make its own models the default choice.
Running "model + Harness" as twin engines on the same stage, the business logic is clear: the model sets the ceiling for intelligence, the Harness shapes its form — and the richer the forms, the more scenarios burn tokens.
70,000 stars in a day shows the community has real demand for an "assemblable open-source foundation." Whether that demand becomes a lasting ecosystem depends on how the plugin market and breaking-change management look six months from now.
Want to try it? Node.js installed, one npx command, and you're off.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
