Blog · Guide
How to pick DSH plugins: read the plugin tree first, then choose by task
September 01, 2026 · dshbase · rewritten field note
Rewritten from a WeChat article (account: 硅基直觉) on the DeepSeek Harness plugin ecosystem; plugin directions are kept as stated in the source and are not a dshbase rating.
Source clue (WeChat via Sogou; not a verbatim reprint) — Rewritten from a WeChat article (account: 硅基直觉) on the DeepSeek Harness plugin ecosystem; plugin directions are kept as stated in the source and are not a dshbase rating.
Scroll the DeepSeek Harness topic page and the first impression is noise: hundreds of repos, a lively but messy market. The number of plugins is not what deserves attention. The organizing logic behind them is.
The core is not bolting tools onto a model; it is splitting the whole agent runtime into replaceable, re-stackable plugins. Model, tools, skills, sessions, sandbox, storage, loops, scheduling and UI are not baked into the kernel. They are assembled from plugins. The official phrase is blunt: every capability is a plugin.
Why build the system as a plugin tree
Think of an agent as a car: the model is only the engine. The steering wheel, brakes, navigation, dashboard and safety rails all have their place. Harness takes those parts apart, so a running dsh is a plugin tree: profile, bundle and patch layers stack in order to produce the capability set a session sees.
The design pays off in three ways:
- Replaceable: switch to another model adapter tomorrow without touching the workflow.
- Isolated: risky operations go through the sandbox and approval layer, so permissions do not scatter across every tool.
- Composable: one capability can be mounted as separate tool, event, UI, storage and policy layers.
Six plugin families: learn the boundaries first
Before memorizing repo names, learn the capability boundaries:
- Model: connects different LLMs or multimodal models; useful when you mix providers or switch models.
- Tools: let the model actually do things — search the web, read and write files, call external systems, run commands. The agent's hands and feet.
- Session and storage: record context, persist state, replay history, resume tasks. Essential if you do not want the agent to forget everything each turn.
- Sandbox and approval: constrain high-risk actions, limit the execution environment, decide what needs human confirmation. The closer you get to real workflows, the more this layer matters.
- Orchestration: loops, scheduling, subagents, background tasks. This decides whether the agent only answers questions or actually drives work forward.
- UI: web UI, desktop shells, panels, message cards. For many people the entry point decides whether the system gets used daily.
Directions worth trying first
If you only want a few directions to open, match them to your work:
- Visual output — prototypes, boards, pages, slides, images:
nexu-io/open-design - Architecture, flow and data-flow diagrams:
tt-a1i/archify - Long-term memory across days, weeks and projects:
MemTensor/MemOS,EverMind-AI/EverOS - Terminal coding agent:
esengine/DeepSeek-Reasonix - Multi-agent collaboration and swarm:
ruvnet/ruflo - Convenient desktop entry:
anywhere-labs/dsh-desktop - Image uploads, external links and asset management:
Molunerfinn/PicGo - Directory repos like awesome-dsh-plugin: good as an entry map, not as your first installs
How to choose: by task, not by hype
One principle covers the whole selection game: try what directly strengthens your current workflow first; set aside what only looks lively. In a plugin ecosystem the real cost is not installation. It is attention spent on directions that do not fit.
Before installing random GitHub plugins on dshbase, check the verified directory and the audit notes.
New to DSH? Run one small experiment: pick the plugin direction closest to your work and ask whether it removes a step or merely adds a button. The former deserves a try; the latter deserves a bookmark.