Sep 22, 2026 · 10 min read · AI Agents

DeepSeek Harness Explained: Install, Modes, Plugins & Claude Code Comparison

On 13 August 2026, DeepSeek open-sourced DeepSeek Harness (dsh) under MIT — and developers piled in by the hundreds of thousands. The pitch is simple: a model is the soul of an agent, but the harness — tools, sessions, sandboxes, permissions, the loop — is what makes it work. DeepSeek's bet: make every piece of that harness a swappable plugin. This is the full guide: architecture, modes, install steps, use cases and the honest caveats.

What a harness actually is

A language model on its own cannot touch your files, run tests or remember yesterday. The harness is everything around it: file editing, shell access, web search, session history, approval policies, subagents, sandboxes. Same model + better harness = better agent. That is why DeepSeek shipped the harness as its own product instead of just a better model.

Everything-is-a-plugin architecture

Most harnesses have a fixed core with extension points bolted on. DeepSeek Harness inverts it: the model adapter, tool registry, session log, agent loop, sandbox and even UI are all Cordis plugins. Cordis is the tiny meta-framework underneath — it only mounts, unmounts and wires dependencies. Consequence: want a different model, sandbox or loop? Mount a plugin beside the others. No fork, no patching a privileged core.

Observability built in

Every run writes an append-only session log — system prompts, reasoning, tool calls, results, subagent scheduling, every context injection. The Trajectory view lets you inspect it by source; resume, fork, search and replay all run on the same event stream. When agents fail, you get traces, not guesswork.

The 4 modes explained

Modes are presets, not silos — Creator mode exists so you can compose new ones from existing plugins.

Install + first run in 5 steps

Step 1 — Prereqs

Node.js ^22.19.0 or 24+, npm/npx, a DeepSeek API key with balance, a disposable test project.

Step 2 — Launch

npx @deepseek-ai/dsh web

Open http://127.0.0.1:3080 (use --port if busy).

Step 3 — Add model key

Settings → Models → DeepSeek provider → paste API key → save. Start with deepseek-v4-flash (cheap); switch to Pro for hard tasks. New sessions pick up model changes.

Step 4 — Workspace + permissions

Select your project folder as workspace (composer stays disabled until you do). Choose the workspace-write preset — confinement plus approval prompts. Avoid danger-full-access on real machines.

Step 5 — Give it a real task

Inspect this repo. Run npm test, find the root cause
of the failing test and fix it. Smallest change only.
Re-run tests, show the diff, report risks.

Where to use it (use cases)

• Agentic coding: multi-file refactors, test-driven bug fixes with V4 Pro
• CI/headless jobs: one-shot runs in pipelines with the headless profile
• Custom tooling: mount MCP servers, company gateways, memory plugins
• Model benchmarking: Minimal mode + any provider to compare models fairly
• Offline work: point it at Ollama/vLLM via OpenAI-compatible endpoints

Where NOT to use it (yet)

Production-critical pipelines — it is an explicit developer preview with breaking changes warned. Keep Claude Code/Codex for production; evaluate dsh beside them.

DeepSeek Harness vs Claude Code

Notable: dsh can even call Claude Code or Codex as subagents — less a rival, more a chassis that can sit above them.

Pricing

Harness itself: ₹0, forever (MIT). You pay model usage — DeepSeek V4 Flash/Pro API rates, with off-peak discounts and cache-hit savings up to 90%+. Note DeepSeek raised API prices in August 2026 with peak/off-peak tiers, so schedule heavy runs off-peak and keep cache prefixes stable.