DeepSeek Harness Explained 2026: What It Can Do, How It Differs from Codex & Claude Code, and the Best Way to Use It

"DeepSeek V4 launched together with 'DeepSeek Harness' — but what is it? How is it different from Codex and Claude Code? Which should I use?"

Bottom line first:

DeepSeek Harness (dsh) is an AI coding-agent "harness" that DeepSeek released on August 13, 2026 under the MIT license (open source). Its biggest feature is "Everything is a plugin" — models, tools, sandboxes, sessions, and even the UI can all be swapped and recomposed freely. Unlike Claude Code (closed) or Codex (closed product), it is fully open source and works with models other than DeepSeek (Claude, GPT, local, etc.).

This article is a beginner-friendly re-organization based mainly on the official page “DeepSeek Harness developer preview” and the GitHub repository deepseek-ai/deepseek-harness, with reference to several reviews and comparisons.

This article does not contain affiliate links.

What is DeepSeek Harness: “Everything is a plugin”

DeepSeek Harness (project name dsh) is an agent harness released by DeepSeek — the "harness" concept explained in our previous article: the software environment that lets an AI model work as an agent.

What sets it apart: everything is a plugin.

Component	Role
Model	Connect any API: DeepSeek, Claude, GPT, local...
Tools	File editing, shell, search, web...
Sandbox	Execution safety configured via plugins
Session	Records everything it sees; search & replay
Agentic loop	The working loop (swappable)
UI / CLI / SDK	Web UI, command line, Python SDK

The central Cordis kernel only manages plugin mounting, dependency resolution, and event dispatch (no business logic). All features are plugins, recomposable in config (YAML) without touching source code.

How DeepSeek Harness works: around the Cordis kernel, models, tools, sandbox, sessions, loops, and UI are all plugins

DeepSeek Harness is not a "packed product" but a "framework" you can reassemble.

What can DeepSeek Harness do?

Like other coding agents, it has the core capabilities, plus its own strengths.

Core capabilities

Repository inspection, file editing, shell/command execution, web/file search
Planning, delegating to subagents
Skills (packaged workflows), goals, multi-step workflows

Unique strengths

Model-agnostic: officially 40+ model providers (OpenAI, Anthropic, Google, Kimi, etc.) plus any OpenAI-compatible endpoint (self-hosted, local GGUF)
Fully traceable: everything the model sees is recorded in an append-only session log. Inspect by source in the Trajectory view; Resume, Fork, search, and Replay all work on the same event stream
Fully open source (MIT): read, fork, embed, and modify the source. Local-first: models, credentials, and sessions stay on your machine
Multiple runtime modes:
- Standard mode: full coding agent
- Code mode: model-generated code (Code Mode SDK) orchestrates multi-step tool calls
- Minimal mode: minimal (bash + file editor) for quick model benchmarking
- Creator mode: build your own custom agent presets

How it differs from Codex & Claude Code: 3-way comparison

Here's a comparison on the key axes (based on public info as of the v0.1 preview):

Axis	DeepSeek Harness	Claude Code	Codex
Open source	Fully (MIT)	Closed	CLI open, cloud managed
Model lock-in	None (any model)	Claude-focused	OpenAI-focused
Extensibility	Everything is a plugin	Hooks/MCP/plugins	Limited
Sandbox	Plugin-defined	App-layer hooks	Kernel-level (strong)
Cost	Model API only (cheap with V4-Flash)	Subscription or API	Subscription or API
Maturity	Developer preview	Stable product	Stable product

Key takeaways:

Open source & auditability: Only DeepSeek Harness opens its full source, lets you audit what the model sees and what tools touch, and runs fully offline locally
Model freedom: DSH is model-agnostic (DeepSeek default, but Claude, GPT, local, or multiple providers can be routed in one session). Claude Code is Claude-centric; Codex is OpenAI-centric
Extension philosophy: vertical integration (Claude Code/Codex) vs horizontal composition (DSH). DSH sells "you can build the engine yourself"
Sandbox strength: Codex's kernel-level sandbox is strongest for isolating untrusted external code; DSH lets you build your own policy via plugins
Cost: DSH itself is free. With DeepSeek V4-Flash, large-scale automation can be cheaper than incumbent subscriptions
Maturity: DSH is a developer preview — breaking changes are possible right now. Claude Code and Codex are stable products

How to use it best: recommendations by use case

"Which to pick" depends on your goal. Here's my honest take:

Goal / situation	Pick	Why
Cost-sensitive, large-scale AI automation	DSH + DeepSeek V4-Flash	Free harness; pay only model API, cheap
Full customization, own policy, audit/compliance	DSH (open source)	Source audit, plugin sandbox, local run
Switch between multiple models (Claude/GPT/local)	DSH	Model-agnostic routing
Stable, ready-made product in a big ecosystem	Claude Code / Codex	Mature, supported, out-of-the-box
Strong isolation for untrusted external code	Codex	Kernel-level sandbox is strongest
Complex multi-stage approval workflows	Claude Code	Fine-grained permission modes, enterprise controls

Notes

DeepSeek Harness is a v0.1 developer preview. The official docs say there will be breaking changes. For mission-critical production, Claude Code/Codex are safer today
DSH assumes "you build it yourself," so setup is harder than Claude Code/Codex. Best for engineers who can follow the source and enjoy experimenting
Many developers realistically run both: DSH for experiments (cost/freedom) and Claude Code/Codex for production

Summary: DeepSeek Harness is the open-source harness for those who choose "maximum freedom"

DeepSeek Harness is a fully open-source (MIT) AI coding harness built on "everything is a plugin" — models, tools, sandboxes, sessions, and UI can all be recomposed freely. It is model-agnostic, and with DeepSeek V4-Flash it keeps costs low. However, it's a developer preview (v0.1) — for stable production use, Claude Code/Codex; for freedom and auditability, DSH. That split is the optimal way to use them.

If you value cost, open source, and free customization, DeepSeek Harness is a choice worth watching in late 2026.

Sources

DeepSeek official「DeepSeek Harness developer preview: Everything is a plugin」(published Aug 2026)
GitHub「deepseek-ai/deepseek-harness」(MIT license, v0.1 preview)
Comparison/review info based on public articles as of the v0.1 preview; figures refer to the primary sources above

※ This article is a beginner-friendly re-organization/paraphrase of the primary sources; some parts are the author's supplementary explanations (based on public information, not speculation).