Published on the dev-workshop channel · English
On August 13, 2026, DeepSeek did something unusual for a model lab: it open-sourced the harness — the layer that sits between an AI model and the real world — and named it DeepSeek Harness, or simply dsh. The tagline on the repo says it all: "Everything is a Plugin."
Within hours, the project had crossed 80,000 GitHub stars. It now sits near 172,000 stars with more than 18,000 forks and roughly 12,900 commits. More striking than the star count is what exploded alongside it: an ecosystem of community plugins numbering in the thousands barely a week after launch. This is the story of what dsh actually is, what its ecosystem looks like, what DeepSeek is trying to build, and how it stacks up against the competition.
DeepSeek Harness is DeepSeek AI's open-source, MIT-licensed agent harness. To understand it, you have to separate the word "harness" from "model." A harness is the scaffolding around a model: it assembles prompts, exposes tools, records what the model does, and decides what the model is allowed to touch. dsh is DeepSeek's own implementation of that layer.
The defining architectural move is spelled out in four words: everything is a plugin. In almost every other agent framework, there is a fixed core that you extend around. In dsh, there is no privileged core to patch. The model adapter, tool registry, session store, sandbox, approval policy, and even the UI all mount through the same plugin kernel — built on Cordis, a plugin framework DeepSeek had been developing for years (created May 2022, MIT licensed), whose design is described in the paper A Programming Paradigm for Spatiotemporal Composability.
Agent = Model + Harness. The model is the soul; the harness keeps the agent working in real-world environments — understanding its environment and using tools.
There is deliberately no interactive CLI. dsh boots as a local web app served at http://127.0.0.1:3080 by default:
npx @deepseek-ai/dsh web
Or from source:
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web
That terminal-less design was a choice, not an omission — a TUI package was built and then deleted the week before launch. The community promptly restored a terminal interface anyway (projects like oh-dsh), which tells you how much the choice mattered to some developers.
str_replace_editor), used for benchmarking models.The feature that made developers sit up: an append-only session event log. Everything the model sees is recorded — system prompts, reasoning blocks, tool calls and their results, subagent scheduling, every context injection. In the Trajectory view, you can inspect these records by source, then resume, fork, search, or replay back along the event stream. When an agent derails at step 40, you fork at step 39 and retry — with full visibility into what it saw.
The ecosystem grew with a ferocious velocity. The dsh-plugin GitHub topic carried ~890 public repositories one day after launch. By the time the Chinese article this blog responds to was published (August 15), the community counted 911 plugins. It has grown well past that now.
The plugin catalog is wild and already diverse:
Memory & context: dsh-simple-memory, dsh-hme (cross-session long-term memory), dsh-memory-vault, context managers like context-vista and billion-context-dsh (model-driven context compression).
Agents & orchestration: dsh-plan-execute (dual-model plan/execute routing), dsh-collaboration (multi-agent squads), dsh-todo-guard, dsh-agent-arena (compare coding agents in isolated Git worktrees), dsh-self-evolving (evidence-first self-evolution engine).
Search & models: dsh-github-intelligence, dsh-hacker-news, dsh-web-search, vision bridges for text-only models.
UI & IDE: DSH-better-sidebar (turns the Web UI into a full IDE), dsh-open-in-vscode, themes.
Fun: desktop pets, games, and lifestyle plugins.
Four separate curated index/awesome-list efforts launched within days of the project going public — awesome-deepseek-harness, awesome-dsh-plugin, and handbook projects like the "DSH orange-book." That ratio (many guides, few stable plugins) signals hype racing ahead of maturity — but it also signals that developers expect this ecosystem to matter.
The Chinese article that kicked this off — 五照实用插件 / "5 practical DSH plugins tested" — showcased a useful cross-section of the ecosystem. Here's the run-down:
dsh-plugin-check — the plugin health checkWith plugins arriving that fast, you need a hygiene tool. This one scans a plugin's manifest, patch, build artifacts, profile bundle, and Hub listing — 33 checks total — and reports pass, warn, or fail. It's read-only and never executes a build. Perfect for vetting a community plugin before installing.
dsh plugin --profile web add github:omdsh-dev/dsh-plugin-check
And to remove a conflicting plugin that breaks startup:
npx @deepseek-ai/dsh plugin --profile web remove <package>
dsh-im-bridge / WeChat bridge — control dsh from your phoneThis bridges DeepSeek Harness into WeChat, so you can send tasks, check status, switch sessions, and approve human-in-the-loop actions from your messenger. Type /yes or /no to continue an agent awaiting approval.
dsh plugin --profile web add https://github.com/BiBoyang/dsh-im-bridge
# also add to headless profile if you run headless
dsh plugin --profile headless add https://github.com/BiBoyang/dsh-im-bridge
Caveat: v0.1 uses WeChat's unofficial iLink gateway, so there are account-risk warnings — fine for experimentation.
whale-girl — the desktop petPure delightful: a desktop pet that lives in the corner of the DSH Web UI. Drag it, feed it, play with it. The pet mirrors the agent's mood — it thinks while the agent thinks, celebrates on task completion, gets sad on failure, and sleeps when idle. Everything comes together as levels, titles, and memories.
dsh plugin --profile web add "github:vlln/whale-girl#main"
dsh-open-in-vscode — one-click handoffSatisfyingly simple: adds an Open in VSCode item to the Workspace menu, so you can jump straight from the agent's project into your editor.
DSH-better-sidebar — turn dsh into a real IDEFeeling the Web UI is too minimalist? This adds a right sidebar + bottom panel dual workbench: file explorer, file preview, embedded web browser, real terminal, Git Diff, background tasks, and sub-agent status — all in one page.
# macOS / Linux (or Git Bash / WSL on Windows):
curl -fsSL https://raw.githubusercontent.com/omdsh-dev/DSH-better-sidebar/main/scripts/install.sh | bash
# Windows (PowerShell):
irm https://raw.githubusercontent.com/omdsh-dev/DSH-better-sidebar/main/scripts/install.ps1 | iex
Read the vision chargelessly and it's clear dsh isn't trying to win a "best coding agent" popularity contest — it's trying to re-architect the agent runtime itself.
No privileged core. DeepSeek's bet is that an agent harness should be composable at every level. Swap the model adapter, the session store, the sandbox, even the UI — all as configuration. The README says it plainly: there is no privileged core to work around, you extend dsh by mounting a plugin beside the others, never by forking a center.
Provider-agnostic by design. Only two model adapters ship today (llm-deepseek, llm-pi-ai), but the adapter contract is documented so any OpenAI-compatible or custom endpoint can plug in via configuration. In theory it's the least DeepSeek-locked of DeepSeek's tools.
Traceability as a first-class feature. The append-only session log makes every run replayable and auditable. That's a real differentiator — most agents live in terminal scrollback; dsh gives you an event stream you can fork and replay.
The community defines the capability boundary. When a harness lets plugins grow new tools, new workflows, and new agent forms, its limits are set by whoever shows up to build them — not by its vendor. This is deliberately broadcast ("Join the dsh plugin ecosystem"), and it's the thesis that turned a Chinese AI lab into a infrastructure project.
dsh arrived into a crowded, fast-moving category. Here's how the field breaks down:
| Tool | What it is | Key posture vs dsh |
|---|---|---|
| Claude Code (Anthropic) | Proprietary, terminal-native coding agent | The closest closed comparison — dsh is positioned as its open-source rival; also can delegate to Claude Code as a subagent |
| OpenCode (SST/Anomaly) | Leading open-source coding agent, ~160k stars, millions of daily users | The "safer daily driver": mature TUI + IDE + 75+ providers; but it extends a fixed core, it can't replace it |
| OpenAI Codex | OpenAI's coding agent | Competitive reference point; dsh can also bridge to it as a subagent |
| Amazon Kiro | AWS agent-centric IDE | Part of the broader agent-IDEC workspace race |
| AWS Bedrock / Google Vertex / Antigravity | Enterprise agent platforms | Cloud-side competition for managed agent hosting |
OpenCode is the safer daily driver; DeepSeek Harness is the more consequential architecture.
The honest caveat: nobody has credible head-to-head coding benchmarks yet. dsh is days old, and some early "DSH benchmarking" online actually tested unrelated community wrappers with similar names. If the categories matter to you, run your own controlled trial: same model, same endpoint, fresh identical repo clones, same task, three runs minimum, scored on correctness, diff quality, tests passed, wall-clock time, tokens, retries, and cleanup needed.
DeepSeek Harness is a developer-preview project — version 0.1.0-rc, with the README promising compatibility-breaking changes. It is exactly the kind of thing you explore, not something to depend on for production deadlines today. But architecturally, it's one of the boldest statements about what an agent runtime should be: fully composable, fully traceable, provider-agnostic, and community-defined at the edges.
Whether dsh becomes the WordPress of agents or just a fascinating construction that dies under its own churn, the open question it poses is worth sitting with: if the harness has no privileged core, then what can't the community rebuild next?.
This article was researched and written in response to the article "五款实用 DeepSeek Harness 插件" and published to the dev-workshop channel. Sources include the official deepseek.com/harness page, the deepseek-ai/deepseek-harness repo, the awesome-deepseek-harness catalog, and comparable-project analyses.