17.1 What Is DeepSeek Harness: Cordis Kernel and "Everything Is a Plugin"
🐋 "Model + Harness = Agent; the model thinks, the harness acts."
One-Sentence Definition
DeepSeek Harness (dsh / DSH) is an Agent runtime framework open-sourced by DeepSeek AI on August 13, 2026 under the MIT license. It is not a model, not an API client — it is an AI operating-system shell.
Why DeepSeek Built an Open Harness
By late 2025, the action-Agent race had two clear poles: commercial closed harnesses (Claude Code, OpenAI Codex, Cursor — mostly model-bound) and open consumer harnesses (OpenClaw, Hermes — model-agnostic). DeepSeek chose a third path — the workshop route: decouple "capability" from "model" completely, ship 100+ default plugins, and four run modes, all MIT.
The team's launch slogans: "Everything is a plugin" and "Closed source gives users the present; open source lets users own the future."
What "Everything Is a Plugin" Means
DeepSeek Harness runs on Cordis (a plugin meta-framework by the Koishi team) — a microkernel that only loads/unloads plugins and manages their dependencies. On top of it, every Agent capability is a plugin:
| Agent capability | Cordis plugin | Default implementation |
|---|---|---|
| LLM call | llm-* | anthropic / openai / openai-compatible / ollama |
| Tool execution | tool-* | shell / fs / edit |
| Skill system | skill-* | Markdown loader |
| Context | context-* | summary / window |
| Sub-agent | subagent-* | spawn |
| Session | session-* | sqlite |
| Storage | storage-* | local / redis |
| Sandbox | sandbox-* | local / docker |
| Scheduler | scheduler-* | cron |
| UI | tui-* / web-* | ink / react |
100+ default plugins — that's what "everything is a plugin" really means.
Three direct benefits
- Swap models without touching the harness — one config line;
- Add capabilities without touching the core — write a plugin, enable, restart;
- Remove/disable features easily — disable
web-*,tool.shell, etc.
📌 vs the previous two chapters: OpenClaw makes capabilities into Skills/plugins but keeps a privileged core; Hermes does the same but has a built-in loop; DeepSeek Harness goes further — even the Agent loop itself is a plugin.
Core Feature List (per main branch)
| Feature | Note |
|---|---|
| MIT open source | full source usable commercially |
| Everything is a plugin | Cordis microkernel + 100+ default plugins |
| Model-agnostic | any OpenAI-compatible endpoint + custom adapters |
| Four modes | standard / minimal / PTC / create |
| Event stream + trajectory replay | git-style replay / fork |
| MCP compatible | built-in MCP server |
| SKILL.md compatible | interoperable with Claude Code / OpenClaw / Hermes |
| Local + sandbox | default local, docker/kubernetes options |
| CLI + Web | dsh + browser at 127.0.0.1:3080 |
| Python SDK | import deepseek_harness |
Verifiable Facts
Repository deepseek-ai/deepseek-harness; MIT; TypeScript core + Python SDK; Cordis upstream koishijs/koishi; default install npx @deepseek-ai/dsh web; docs deepseek-harness.github.io/deepseek-harness/; ~40 swappable models; MCP + Anthropic SKILL.md compatible; associated paper "A Programming Paradigm for Spatiotemporal Composability"; released 2026-08-13 (v0.1 preview). Time-varying numbers are not cited.
Spectrum Position
Action-oriented Harness spectrum (2026)
├── multi-channel · self-evolving (open) → Hermes (Ch.15)
├── multi-channel · consumer · skills (open) → OpenClaw (Ch.14)
├── terminal IDE · industrial · closed → Claude Code (Ch.16)
└── terminal+web · workshop · all-plugins (open) → DeepSeek Harness (this chapter)
The true differentiator: only DeepSeek Harness makes the Agent loop itself swappable.
Section Summary
| Topic | Key point |
|---|---|
| Positioning | open Agent runtime, model-agnostic, pluginized |
| Team | DeepSeek AI (lead Tianyi Cui) |
| Upstream | Cordis (Koishi team) |
| Core differentiator | "everything is a plugin" — even the loop |
| Protocol | MCP + Anthropic SKILL.md |
Next section: 17.2 Installation & 4 Run Modes