Plugins

Paste a link; the agent reads it. Xiaohongshu, Douyin, Bilibili, WeChat, articles, PDFs and audio become citable Markdown for a wiki or knowledge base. On 60 turns with vs without: finishes reading 22% -> 92% (+70 points), times out 68% -> 0%, answer quotes a real source fact 23% -> 82% (the A/B is exploratory: different time caps, no human correctness review). On 123 live links 80% came back usable (93% of the pages that answered); an unreachable link reports in 26 s. Codex, Claude Code, DSH.

Counts persistent operations within a turn and escalates on the total: a reminder first, then an explicit stop instruction, then an approval prompt. Reads the session sandbox mode on every call, and stays silent under read-only and workspace-write.

All-in-one companion for the DSH web GUI: a balance whale widget with today's usage, per-turn cost and peak/off-peak pricing, an editable bubble queue and sound effects; a 106-clip transparent desktop pet with spring drag, throw, screen roaming and clones; a QQ message card bridged through SnowLuma for reading history and replying; Bilibili dynamics and comment interaction with a fish guard and a day/night schedule; and local, redacted two-way memory. One plugin; data stays on the machine. Runs on Windows 10/11 only.

A Live2D whale-girl companion for the DeepSeek Harness Web UI: the model stays on screen and reacts to what the agent is actually doing (thinking, tool calls, streaming text, errors, turn completion); click her to send the agent a message and watch the reply stream into a speech bubble; right-click for a menu of expressions, dress-up items, desk scenes and one-shot animations, every action mapped from the original model author hotkey sheet so nothing overlaps and everything expires on its own. Non-commercial model assets (CC BY-NC-SA 4.0), MIT code, by Andersen216 (https://github.com/Andersen216).

边做边压 Live Compact. Compress the finished turn in place and keep the prefix cache.

Desktop notifications for turn completions, with per-outcome controls and keyword rules.

Turns DeepSeek Harness into a server-grade multi-tenant platform: remote access + auto HTTPS, subuser permissions & token/daily quotas, sandbox enforcement, encrypted auth & audit log.

Reversible context slimming for DeepSeek Harness: TOON-style lossless tabular re-encoding of uniform JSON arrays, stride-sampled table windows with line numbers, and cross-turn dedup for repeated tool output; every replacement passes a byte-plus-token gate (≤72% bytes, strictly fewer estimated tokens) and the original is spilled to disk behind a `[save-token #id]` notice retrievable in one call via the `save_token_expand` tool, with billed and avoided token metering on a settings dashboard.

Offline meme and slang dictionary: 6,792 built-in entries are looked up before a turn and injected only when a message looks like a meme question. Ships a lookup tool, a settings page for adding/editing/deleting entries and aliases, an adjustable tone slider for how far it plays along, and a built-in version check.

AstrBot/DSH-native GPU coordination for local AI companions: LLM, IndexTTS and ComfyUI take turns, no proxy.

Request-side injection and jailbreak plugin for DeepSeek Harness: rewrites only the request copy sent to the model — system-prompt sections, fabricated turns, and keyword or regex replacement — leaving the conversation log and host files untouched; ships unrestricted / nsfw / authority presets, imports the dsh-purge 破甲 prompt as an ordinary preset, enables per agent preset, always-on, or by model and agent preset, and records each injection as a collapsible row in the session.

Log query card for the DSH Web GUI over Elasticsearch / OpenSearch: one query surface with per-dialect query strings and clickable examples, field-name completion, filter chips, a time histogram whose bars zoom into that window, field facets that turn a bucket into a filter, and a result list with level colors, stack-continuation rows, columns and paging; Docker container logs are a collection source that reuses the docker plugin. Ships five read-only loglens_* tools (sources / search / fields / facets / context) sharing the same execution layer as the panel. Outbound requests only reach configured sources, judged on the resolved IP, without following redirects and with size and timeout caps.

IM webhook and local notifications on turn completion, errors, or approval (Feishu/WeCom/DingTalk/Slack/Discord/custom).

Design workbench in the right sidebar. Its tools sit in floating groups at the canvas corners and a bottom dock with text labels, and a home screen offers one prompt box, recent designs and large template cards. When the home screen's style is Auto, the plugin picks a style card that fits the subject and hands its colors and fonts to the model; your own words, a chosen card, a reference image or a default brand design system take precedence. Each design gets its own folder. When a design needs photos, the model can copy from a library of 64 AI-generated placeholder images in 8 themes (about 8 MB, downloaded from npm the first time they are needed; made with OpenAI image generation and screened by a person, not real places, products or people); export and handoff remind you to replace them before launch. When the model builds web pages, app screens, desktop app windows, WeChat mini programs, posters, slides or emails, each HTML file shows up as a screen on a canvas inside a device frame and refreshes when the file changes; screens far from the view show a background thumbnail and load the live page when you move closer; while the model writes a page, the sidebar grows with it and a few cursors draw boxes, fill colors and type where it is writing, with a timer in the status bar; local dev servers are proxied with hot reload intact, and an existing HTML file or folder opens by browsing or by dragging it into the sidebar (desktop app). A design library gathers every design you made in any session, your saved inspiration (reference images, URLs, style cards) and a design market of shared templates, design systems and style cards (16 templates, 3 design systems and 20 style cards to start; search, filters, full-size screens and a live view; an offline copy ships with the plugin): start from one in a click, or share your own design to the market after a personal-data scan and an offline design check; after connecting GitHub once (public repositories only), submitting takes one click, and the web form still works. Mini program screens get the WeChat capsule in the top-right corner, which turns white over a dark top bar, and the design check flags text and buttons hidden under it. App and mini program screens follow platform layout (status and nav bars, a tab bar of 2 to 5 items, safe areas), desktop app screens get macOS traffic lights or Windows caption buttons, and the design check flags content under the Dynamic Island, the home indicator or the window buttons. Click an element or drag a region to put its selector, source file and styles into the composer, or draw on the preview (pen, highlighter, arrow, box, circle, text) and hand the marked-up screenshot to the model; picks, markup and style changes appear in the composer as small chips while the model still gets the full text. It can extract a brand design system from your website or project (colors by role, fonts, type scale, radius, shadows, buttons, inputs and cards) and use it for every new design. The model adds a few tweak sliders to each design (spacing, radius, brand color, layout variants) that update all screens live and write back to the stylesheet in one click. A finished design can be turned into React, Vue, Next.js or plain HTML code, or for mini programs into uni-app or native WeChat code; existing uni-app (including HBuilderX projects without package.json), Taro and native mini program projects are recognized so new pages go straight in, and each built screen is diffed pixel by pixel against the design. Also has a design tokens panel, twenty style presets (each a live sample screen you can open full size), a style panel (font size, colors, radius, padding and more) that writes back to the HTML or hands the change to the model together with the CSS rule and line behind each property, versions saved before and after each turn with drag-to-compare and restore, export to images, PDF and PPTX (links between screens stay clickable), a LAN QR link for viewing on a phone, link arrows and a click-through try-it mode, a multi-size view, in-place text editing, reference-image palettes, and a design check (overflowing, clipped or overlapping text, contrast, tap target size and more). Bundles a ui-design skill and seven tools: preview_page opens a preview; check_design measures every screen at real sizes in a hidden local Chrome or Edge (bars covering content, script errors, missing assets, dead links; CSS file and line for contrast and font-size issues; cropped screenshots of serious problems); look_design hands real screenshots to the model; extract_design_system builds a design system from a site or project; handoff_design writes a handoff pack (screenshots, routes, tokens, shared components); compare_design diffs the built local page against the design and marks what differs. find_photos copies matching placeholder photos from the photo library into the design. Adds a Design mode for new sessions (Standard mode's tools plus the skill, the seven tools and a design persona); each part has its own switch on the Plugins page. When DSH is opened from a phone or another computer, local-file designs are bundled into a sandboxed frame so they still show, and unchanged screens are not sent again. An experimental option, off by default, has the model also write business annotations (rules, states and edge cases the UI does not show), shown as numbered markers on the page and cards in the sidebar and passed on in code handoff.

Codex-style workflow auto-collapse: finished turns fold into a single "processed in Xs" row leaving only the final answer; tool calls and think blocks collapse into live-summary chips that follow streaming; expand level by level, fully reversible on uninstall.

Per-model working languages: ordered regex rules map model ids to BCP 47 language codes, your prompts are translated into the model's language with DeepSeek Flash before sending and replies are translated back after each turn, with in-place toggles for the original reply and the text actually sent, a system-prompt rule that keeps bridged models in their working language, a warning under the message when translation fails, and no translation or cost when the languages already match.

Lark/Feishu bot channel for DeepSeek Harness: each chat drives its own agent, and tool approvals, model questions, and plan reviews return as cards answered by a button or a reply. Switch workspace and model from the chat (`/cd`, `/model`, `/new`), and run several bots that keep separate sessions and can hand turns to each other in one group.

A Claude-persona desktop pet pinned to the bottom-right corner of the DSH web GUI — shows balance, today's usage, peak/off-peak pricing and per-turn cost, with editable bubbles (text, balance, usage, pricing, images, random lines and weighted choices), per-line styles and fonts, a floating quick editor, sound effects, and custom character/audio. Ships animated roles that breathe and blink, plus a balance-alert variant that switches to a crying face when the balance falls below the alert line; 256/384 animation tiers and an optional 512 HD pack installed from the resource manager. 63 Claude-flavored lines. A derivative of MeteorNOX/DeepSeek-Balance-Whale-Widget (MIT).

Web access for DSH agents: fetch JS-rendered or protected pages as clean markdown, aggregated search across 14 engines, and stateful interactive sessions (click, type, scroll, eval, screenshot) against a hosted or self-hosted Rust browser engine. Every session returns a live view URL so a human can watch the agent browse and take over with the mouse.

Personal assistant for DeepSeek Harness that turns conversational commitments into cross-session plans, reminders, and an auditable dashboard.

Pins the model's reasoning language to a tier picked in Settings - General - Thinking language (Simplified Chinese, or a default tier that only constrains replies and leaves reasoning to the model), by injecting a system-prompt section at order -1000000 and re-hoisting it to first place at the end of the assemble waterfall, plus an optional per-turn copy placed ahead of each user message for models that follow static sections weakly.

Long-context preparation, Dream memory, task verification, workflows and computer interaction for DSH. Needs DSH 0.2.2-rc.2 and Node.js 24.9+. The default install configuration grants full file and command access and turns execution approval off; explicit settings in an existing profile take precedence.

Floating panel for the DSH web UI managing MCP servers and skills: per-server deferred tool loading (ToolSearch, off by default) keeps a server's tools out of every turn's context — the prompt lists tool names only, and a tool's full schema enters the list once the model finds it with the built-in tool_search, so a turn stops paying for schemas it never calls. MCP servers live in a global scope or a project's own .dsh/mcp.json, with per-server enable/disable, a 47-service market and import from ten other clients; skills are auto-discovered across their real scopes with shadowing, and model auto-invocation versus manual-only is a switch per skill.

Installs four multi-agent math-research presets for DSH: v2 (probability-driven pipeline with multi-verifier debate), v3 (paper-style Markdown knowledge base with a planner agent and a reusable method library), v4 (persistent self-organizing residents that message and meet), and v5 (a research institute with an academician who decomposes and assigns work, voting researchers, temp workers, group chat and a compare-and-set task board); all four support checkpoint resume, human intervention, and an optional Lean formal-verification switch (off/encourage/require) whose passing proof turns the vote into a fidelity check of the Lean statements.