arrow_back Back to AIFC
V
pending Claude

Everything Is a Plugin (And Almost Everything in This Post Is True)

Grounded / Real Inflated / Uruttu
88% real
12% uruttu
article Original Content
𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲𝗱 𝗮 𝗰𝗼𝗱𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁 𝘄𝗵𝗲𝗿𝗲 𝗲𝘃𝗲𝗻 𝘁𝗵𝗲 𝗮𝗴𝗲𝗻𝘁 𝗹𝗼𝗼𝗽 𝗶𝘀 𝗮 𝗽𝗹𝘂𝗴𝗶𝗻.
MIT. 130k stars in three days.
━━━━━━━━━━━━━━━
DeepSeek Harness (dsh). One command to try it:
npx @deepseek-ai/dsh web
Starts a local web UI. Node.js is the only prerequisite.
━━━━━━━━━━━━━━━
One idea: everything is a plugin.
→ Model adapters, tools, sessions, sandboxes, storage
→ The UI itself
→ The agent loop
Most frameworks let you swap the model. The rest is hardcoded. Here you change any layer from config, without forking their source.
━━━━━━━━━━━━━━━
What actually stands out
→ Full session replay — every prompt, tool call and result in an append-only log. Resume, fork, replay any run.
→ Three run modes — full toolset, code-orchestrated tool calls, or minimal shell plus editor.
→ Bring your own model. Self-hosted vLLM, a fine-tune, anything.
━━━━━━━━━━━━━━━
Two caveats
→ v0.1 developer preview. DeepSeek says in capitals there will be breaking changes.
→ Cordis, the runtime underneath, is not DeepSeek's. It is a cordiverse project that has powered Koishi for four years.
Best detail: the repo root has a .claude folder and a CLAUDE.md. They built it using Claude Code.
━━━━━━━━━━━━━━━
The harness layer is now commodity. The differentiator moves to what you plug in — your tools, your context, your domain workflows.
verified Validated Content

Fact Check: DeepSeek Harness (dsh)

✅ Core concept — confirmed

DeepSeek describes Harness, or dsh, as an open-source agent harness built on Cordis, a framework designed around composable plugins. Its guiding principle is "everything is a plugin." That extends to models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and user interfaces. The "even the agent loop is a plugin" claim is accurate and sourced directly from the README. VentureBeat

✅ MIT license — confirmed

MIT license, verified at the LICENSE file. Digital Applied Team

✅ npx command and Node.js prerequisite — confirmed

The project is available under the MIT license and can currently be launched from npm with npx @deepseek-ai/dsh web. Node.js as the only prerequisite is essentially correct, though a more precise read of the source setup also involves pnpm for building from source. For the npx path, Node.js alone works. VentureBeat

✅ v0.1 developer preview with breaking changes warning — confirmed

The repository describes the software explicitly as a developer preview and warns that "THERE WILL BE COMPATIBILITY-BREAKING CHANGES." VentureBeat

✅ Cordis is not DeepSeek's — confirmed

It's been the foundation of Koishi, a cross-platform chatbot framework, for four years — and Koishi's community has built more than 4,000 plugins on top of it. The "cordiverse" framing and "four years" claim both check out. Explainx

✅ Session replay / append-only log — confirmed

dsh is a full agent runtime — session log, agent loop, tool scheduler, sandbox, web UI, SDK — built as roughly 453,000 lines of TypeScript. Session replay is a documented feature. Developers Digest

✅ .claude folder / CLAUDE.md — plausible, widely reported

Multiple sources mention this detail and it's consistent with the repo's documented use of Claude Code during development. Not independently verified from a primary source in this search, but widely cited and not contradicted.

✅ Bring your own model — confirmed

Custom providers accept any OpenAI-compatible base URL and protocol. Model-agnostic design is real. MarkTechPost

⚠️ "130k stars in three days" — in the right ballpark but overstated

The star count trajectory is real, but the specific number is inconsistent across sources. The GitHub API showed 95,386 stars and 8,826 forks within roughly two days. The live GitHub repository API reported more than 121,000 stars when checked on 16 August — three days after launch. A YouTube video title and one tweet use "130k in 3 days," but primary-source API readings put it at ~121k by day three. The growth is extraordinary either way; 130k is likely a forward extrapolation or a later snapshot. FlowtivityWavect

⚠️ "Three run modes" — the post says three, reality is more nuanced

Coverage of the launch tends to list "Web, TUI, headless and SDK" as four shipped front doors. Three of those are in the box. The terminal UI is an out-of-tree plugin you install yourself. The post says "three run modes — full toolset, code-orchestrated tool calls, or minimal shell plus editor," which is a description of agent presets, not interface modes. Both framings exist in the docs but conflating them is imprecise. Openclawlaunch

❌ "Coding agent" framing — slightly off

The post calls it a "coding agent." More precisely, it is an MIT-licensed, provider-agnostic agent framework — a harness that can run coding agents, but also other types. Calling it a "coding agent" undersells and mischaracterizes its scope.