Plugins

Fleet console for DeepSeek Harness: one local dsh page manages many remote dsh machines — live idle/running status, in-place switching without new tabs, keep-alive iframes, cross-device session reading, one-click refresh/restart, and a controller/managed role split.

A calmer reading view for DeepSeek Harness with auto-collapsing process, a minimap-style message navigation rail, in-page search, unified failure cards, and generative MCP Apps (SEP-1865) in sandboxed iframes.

DeepSeek Harness plugin: Bilibili keyword video search, video metadata, subtitle transcripts, direct play URLs, and multimodal frame viewing (bilibili_search / bilibili_video / bilibili_subtitles / bilibili_playurl / bilibili_frames). Anonymous by default

Browser-local audio/video toolkit for DSH — extract frames, capture thumbnails, compress images, and convert clips to GIF with FFmpeg.wasm.

Registers the video-studio skill in the DSH skill catalog. It builds a Chinese-narrated video from a script: edge-tts synthesizes the narration, the timeline is derived from each line's measured audio duration rather than estimated from text length, Pillow renders every frame deterministically and pipes the raw frames to ffmpeg as rawvideo, burned-in subtitles and SRT/ASS export come from that same timeline, physical-model sound effects and a generated score are mixed sample by sample with numpy, and QA runs in four layers: container, signal, temporal and page level (the page layer only when the browser engine is used). An optional Playwright-driven browser engine renders the whole film as Canvas2D instead, sending PNG frames through image2pipe. Needs network access for the narration, ffmpeg on PATH and Python with numpy, Pillow and edge-tts; the browser engine also needs Playwright with a local Edge/Chromium.

Drives Windows apps through accessibility controls instead of pixels, with a takeover overlay that frames the window being driven, marks each action with a ghost cursor, and stops on Esc; a browser lane adds sign-in profiles, saved state, and attach-to-my-Chrome. Windows only, and DSH must be run as Administrator for the desktop lane.

Downloads a video link, samples frames from it, and transcribes the speech with Groq Whisper so an agent can answer questions about what the video shows; long audio is split into 15-minute windows to stay under the upload limit, and frames still come back without a key, which only the transcript needs. Needs ffmpeg, ffprobe and yt-dlp on PATH.

A conversation-view tab for turning one local reference video into a human-curated shot list. ffmpeg scdet finds shot boundaries, each shot gets one timecoded thumbnail, you untick what you do not need, and the export writes frames.csv (UTF-8 BOM), shots.json, numbered JPEG frames and a labelled contact sheet. Local-only, no upload, zero runtime dependencies, 39 node --test tests including a real-ffmpeg integration case. Requires ffmpeg 5.1+ (with ffprobe) installed separately.

Pull frames and audio out of a local video with ffmpeg and hand back plain file paths, so the agent can look at the frames with its own vision. Five tools: probe, evenly spaced frames, one frame at a timestamp, scene-change frames (the practical version of frame-by-frame), and audio extraction. No models, no uploads, no routing guesswork.

Serial-port toolkit for DSH - 9 tools plus a web UI where an AI agent and a human share one UART line, with auto port matching, hex/text frames, timestamps, scheduled sends, rule-based auto-recovery and appearance themes. The serial engine runs on Python 3.8+ — DSH's bundled interpreter is used when present, otherwise python / py -3 on PATH.

Runs the local llama.cpp Qwen3.5-4B from a Settings page — Windows only, NVIDIA GPU (8 GB VRAM handles the 4B Q4) — with start, stop and restart, launch parameters, deployment paths with a presence check, provider-route and compaction-policy sync in the profile patch, a loopback request-capture proxy, and a local_worker tool that hands translation, OCR and video-frame chores to the local model by file path. Needs llama-server.exe (verified on b12069) and Qwen3.5-4B GGUF + mmproj files on disk; video frames also need ffmpeg.

Screen frames, log lines and (on Windows) key/mouse input on one millisecond timeline: record, mark the failing seconds, box the problem on the frame, then hand the agent a keyframe strip, an animated GIF and the aligned log slice.

Local video understanding for agents: extracts keyframes adaptively based on motion and pauses, inspects any timestamp at full resolution, and aligns visuals with dialogue from embedded subtitles or configured ASR. Requires ffmpeg and ffprobe; clips without subtitles also need a transcription engine for dialogue (a ~190 MB local whisper model, a cloud API key, or the official voice-input model).

Local searchable archive of DSH chat history: decodes the multi-frame zstd session logs (thousands of concatenated frames), extracts user and assistant text into SQLite, and exposes sync / search / stats tools. Idempotent imports, no external service, nothing leaves the disk.

Video study toolkit for DeepSeek Harness: fetch a public Bilibili single part or YouTube video, sample every 3 seconds plus change-triggered frames, transcribe locally with faster-whisper, page through section grids, and render a 16:9 study deck. Windows only, and needs the DSH pwsh tool plus Python 3.12 x64 with Pillow and FFmpeg/FFprobe; fetching adds yt-dlp, transcription needs faster-whisper with your own local model files, and deck output needs python-pptx.

Wandering Earth 550C control-room palette for the DSH Web UI, with dim 2px frames and amber selection states.

Read a video with a text-only model: with ffmpeg installed, sample frames at a 0.1-10 second interval into contact sheets where every frame carries its real timecode, extract the audio track and send it to iTingnao (needs an ITINGNAO_API_KEY) for a timestamped transcript, and describe each sheet with an image-capable model in the host, so the plugin works on its own without any other vision plugin. A composer slider sets the sampling interval, and pure audio files are transcribed without frame sampling.

A right-Sidebar video player with a folder playlist (first-frame thumbnails, duration badges, sort by name/duration/modified/size, fullscreen, and a context menu for rename/copy/delete/refresh when the folder is opened with write access), live playback telemetry (buffered range, decoded FPS, dropped frames) and a read-only MP4/MOV reader for codec profile/level, exact frame count, per-track bitrate and rotation. Registers no agent tool; file writes use the browser File System Access API only after a write-permission folder choice.

Give a text-only agent video understanding: extract timestamped frame files, sample only on scene changes, export GIF clips, probe metadata, and run a one-shot analyse over the frames. Five tools that shell out to a local ffmpeg through argv arrays rather than a shell string.

Moves the default workspace off the system drive (DSH opens every new Session in <Documents>/deepseek-harness/default-workspace and never relocates it) while keeping stored sessions usable: rewrites the workspace record without a restart, offers the same three actions as a GUI panel, moves stored session logs together with their header cwd, and re-frames logs whose first frame is not exactly one header line. Requires the desktop host (workspaceRegistry and sessionPersistence).

Reads a video or audio file into metadata, sampled stills, subtitles and a multilingual speech transcript with speaker labels, captures the live desktop as frames or a recording, and installs its ffmpeg/whisper.cpp/sherpa-onnx toolchain on first use with per-file sha256 checks; all processing stays local. Automatic toolchain install and desktop capture currently support Windows x64 only, so elsewhere ffmpeg and model paths must be set manually.

Keeps a local sticker dictionary keyed by content hash so an image already described is recalled instead of re-read, and splits a GIF into evenly sampled frames laid out as one contact sheet so the whole animation is visible.

Production pack for ComfyUI inside DeepSeek Harness: submit API-format workflows to a local or remote server, batch-generate and fetch outputs, and diagnose bad images with measured parameter ranges and a failure-mode decision tree (black frames, duplicated subjects, color fringes, OOM).

Read-only lazy session viewer: lists sessions with stat only, then decompresses just the tail Zstandard frames to inspect recent events; includes search, stats, and export in a standalone /lazyview panel. Requires Node.js 24+.