DeepSeek Harness (dsh): Plugins, Cordis Profiles, and the Community Directory (2026)

TL;DR (as of 2026-08-14): DeepSeek Harness (dsh) is DeepSeek’s answer to “model alone is not an agent.” The formula is Agent = Model + Harness — weights plus tools, permissions, memory, and workflow. You extend it through Cordis profiles and patches, install locally with npx @deepseek-ai/dsh web, and browse 488 community plugins in 11 categories on the DeepSeek Harness Plugins directory (synced from the awesome-dsh-plugins GitHub repo). If you have read our Fable 5 vs DeepSeek coding piece, you already know why “harness” matters for benchmarks — this page is about the harness you actually run.

Why “harness” keeps showing up

On fable5.top we use harness in two related senses:

1. Evaluation harness — the scaffold around a model during a benchmark run (tool access, retries, grading). That is why we treat Fable 5’s 80.3% SWE-Bench Pro (Anthropic official) differently from DeepSeek’s self-reported numbers on the same test name: same label, different trust level. See the comparisons hub and the DeepSeek coding comparison for the full split.
2. Agent harness — the scaffold around a model in production: which tools fire, what gets remembered, how sub-agents coordinate.

DeepSeek Harness (dsh) is explicitly the second kind — and DeepSeek ships it as a first-class product, not an afterthought wrapper around chat.

Agent = Model + Harness

A frontier model can pass a coding benchmark and still be awkward in your repo if the harness is wrong: wrong tool budget, no file memory, brittle stop conditions. DeepSeek’s framing matches what we argue in comparison posts — the agent is not the weights alone.

DeepSeek V4’s MIT weights let you self-host the model; dsh is how DeepSeek packages the agent loop that makes those weights useful without rebuilding Claude Code from scratch.

Cordis profiles and patches

dsh extends through Cordis — profile documents that declare capabilities, tool hooks, and behavioral patches. Think of a profile as your agent’s configuration manifest; patches are incremental overlays (a plugin, a team standard, a locale pack) you stack on top.

Typical workflow as of 2026-08-14:

1. Install the base runtime (npx @deepseek-ai/dsh web).
2. Pick a starter Cordis profile (minimal coding agent, research agent, etc.).
3. Apply patches from the community directory or your own repo.
4. Pin versions when a plugin changes tool permissions — patches are powerful because they change what the agent can do, not just what it says.

We have not independently audited every patch in the directory. Treat third-party patches like third-party VS Code extensions: read the manifest, check network permissions, and test in a sandbox repo first.

The community plugin directory

Manually cloning awesome-dsh-plugins and grep-ing README files does not scale once the list passes a few dozen entries. deepseekharnessplugins.com is the browsable front end:

• 488 plugins indexed as of 2026-08-14 (count synced from GitHub; refresh the site if you need today’s number)
• 11 categories — coding, DevOps, browser automation, documentation, MCP bridges, and similar buckets
• GitHub sync from awesome-dsh-plugins so new community submissions surface without waiting for a manual blog roundup

Use the directory to shortlist patches, copy install hints, and compare overlapping plugins before you stack them into one Cordis profile. Duplicate tool hooks are a common foot-gun: two browser plugins fighting for the same permission slot will fail in confusing ways.

Install: npx @deepseek-ai/dsh web

The fastest path to a local UI (verified workflow as of 2026-08-14):

npx @deepseek-ai/dsh web

Requirements in practice:

• Node.js current LTS
• DeepSeek API key (or compatible OpenAI-style endpoint if you proxy)
• A Cordis profile — start from DeepSeek’s defaults, then add patches from the directory

The web command spins up the harness UI locally; it is not a hosted replacement for Claude.ai. Your code and tool calls stay on your machine unless a plugin explicitly calls out to the network — which is why patch review matters.

How this connects to Fable 5 comparisons

We are a Claude Fable 5 reference site, but the harness economics argument is vendor-neutral:

• Benchmark harness parity — only a few SWE-Bench Pro figures share the same evaluation definition. Our comparisons hub calls out which rows are apples-to-apples.
• Production harness choice — Fable 5 users often live inside Claude Code, Cursor, or custom agent stacks. DeepSeek users now have dsh plus a plugin marketplace mindset similar to MCP servers and Cursor rules.
• Cost per solved task — a better harness can make a cheaper model viable; a weaker harness wastes frontier weights. That is the lesson of the DeepSeek coding comparison, applied here to tooling rather than token prices.

If you are choosing models, read the comparison posts first. If you are choosing agent infrastructure on DeepSeek weights, start with dsh, then the plugin directory, then patch deliberately.

Practical checklist

1. Install — npx @deepseek-ai/dsh web and confirm a minimal coding task end-to-end.
2. Browse — deepseekharnessplugins.com for categories that match your stack (CI, browser, docs, MCP).
3. Patch incrementally — one plugin at a time; log tool errors before you add the next overlay.
4. Compare harnesses fairly — when someone cites a benchmark, ask which harness ran it; when someone cites agent productivity, ask which profile and patches were active.
5. Cross-read — Fable 5 vs DeepSeek for coding for model economics; Claude Fable 5 comparisons hub for benchmark harness notes.

FAQ

What is DeepSeek Harness (dsh)? DeepSeek’s agent runtime: Agent = Model + Harness. Cordis profiles and patches extend tools, permissions, and workflow without forking the core loop.

How do I install it? npx @deepseek-ai/dsh web — Node.js, API key, then layer Cordis patches.

Where are the plugins? deepseekharnessplugins.com — 488 plugins, 11 categories, synced from awesome-dsh-plugins on GitHub (count as of 2026-08-14).

Is this the same “harness” as SWE-Bench? Same English word, different layer. Benchmark harness = evaluation scaffold; dsh = production agent scaffold. Our coding comparison explains why that distinction matters for score trust.

Should I switch from Claude Code? Not automatically. Switch when you want DeepSeek’s model economics and an open patchable agent stack. Otherwise, tune the harness you already have and use comparison pages for model routing.

Independent overview as of 2026-08-14. Plugin counts and categories reflect the community directory synced from GitHub; verify on deepseekharnessplugins.com before you plan a rollout. See also: Fable 5 vs DeepSeek coding and the comparisons hub.