The 15 plugins worth installing first

The DSH plugin ecosystem is large and uneven. With over 500 projects indexed across twelve categories, plenty are a weekend experiment, and star counts reward novelty as often as usefulness. This list is filtered differently: which plugins solve a problem you will hit in your first week?

We have organised them by the problem they solve rather than by popularity, because the popular ones cluster heavily in cosmetics.

Category 1: Visibility — knowing what the runtime is doing

Context manager

Agent sessions die by context exhaustion. A context manager shows what is consuming the window — system prompt, conversation, tool output, attachments — and several compress tool results before they reach the model. This is the highest-value install in the entire ecosystem. Every serious DSH setup has one.

What to look for: per-segment breakdown, not just a percentage. Compression that is reversible. A warning before truncation rather than after.

Cost meter

Agent work burns tokens far faster than chat, because every turn carries the whole context. A cost meter shows per-session and per-model spend. Without one the first invoice is a surprise. Several of these are local-first, meaning the numbers never leave your machine.

Session inspector

A step further than a context manager: it shows the actual message sequence the model received, including tool calls and their results. When an agent behaves inexplicably, this is how you find out why. Indispensable for debugging, optional for daily use.

Category 2: Capability — what the agent can do

Vision bridge

DeepSeek's text models cannot see. A vision plugin routes images through a separate multimodal model and feeds back a description, which is enough to debug a UI from a screenshot, read an error dialog, or extract a table from a photographed page. For anyone working with visual material this is transformative; for pure text work, skip it.

Browser control

Headless browser plugins let an agent actually navigate, click and read pages rather than guessing at HTML from a single fetch. Slower and more token-hungry, but the difference between a working task and a confident hallucination. The better ones support a persistent session so the agent does not restart from a blank page each turn.

Document readers

PDF, DOCX and spreadsheet readers that preserve structure rather than flattening everything to text. Table structure is the hard part and the differentiator — most readers lose it, which silently destroys the agent's ability to reason about the data.

Web search

Several free search-provider plugins exist, plus general-purpose search toolkits with more operators. If your agent needs current information, one of these is mandatory; the built-in fetch cannot search. Compare on rate limits and result quality rather than features.

Category 3: Safety — before you give it write access

Command guard

A pre-execution guard intercepts destructive commands — recursive deletes, force pushes, database drops, mass file renames — and requires explicit approval. Install this before giving an agent write access to anything you care about. Retrofitting after an incident is the wrong order.

Sandbox wrapper

Runs tool execution in a container or restricted filesystem so a mistake cannot reach your real files. Heavier than a command guard and worth it for unattended or long-running work.

Category 4: Interface — how you interact

Terminal UI

If you live in a shell, a TUI replaces the default interface with something keyboard-driven and low-overhead. Several exist. The practical difference between them is keybinding philosophy and how much horizontal space they give tool output.

Sidebar framework

Open sidebar frameworks let plugins register their own panels — file tree, terminal, git status, sub-agent views. One install gives you a workspace instead of a chat box. This is the single biggest quality-of-life upgrade for people who work in the GUI all day.

Why we skip themes and pets here

The ecosystem has thousands of stars' worth of cosmetic plugins — themes, desktop pets, animated sidebars. They are genuinely popular and entirely reasonable to install. They just do not change what the runtime can do, so they are not on this list.

Category 5: Operations — running it at scale

Model router

Routing plugins send requests to different providers based on task type, cost or availability. Useful once you have a second provider; essential once you have three. The best ones fail over automatically on rate limits, which turns a hard failure into a small delay.

Memory layer

Memory plugins give agents recall between sessions instead of within one. The mature options are local-first and Markdown-native, which means your memory is a folder you can read, diff and back up rather than an opaque store. Check the storage format before committing — migrating off a proprietary store is painful.

Plugin market

A market inside the application lets you browse and install without leaving it. Convenient early on, when you are exploring. Most people settle into a fixed set within a month and stop using it.

Mobile bridge

Exposes a running DSH session over LAN or a tunnel so you can check on a long task from a phone. Niche, but the only way to supervise an agent that runs for hours.

How to choose without over-installing

The temptation is to install twenty plugins on day one. Resist it. Every plugin adds load time, a version constraint, and another thing that can break on upgrade.

A sequence that works:

1. Day one: context manager and command guard. Nothing else.
2. Week one: add vision if you ever work with images, or browser control if your tasks involve the web.
3. When the bill arrives: add a cost meter and, if you have a second provider, a router.
4. When a session runs long enough to forget: add memory.

Each addition should answer a problem you actually had. The full directory is there when the problem appears — there is no prize for installing early.

On star counts

We rank categories by GitHub stars because it is the only objective signal available at scale. It is not a quality measure. A plugin with 40,000 stars may be a desktop pet; a plugin with 300 stars may be the context manager that saves your sessions. Use stars for discovery, read the README for judgement, and check the last commit date before you depend on anything.