deepseek-harness-studio

Core features

This table lists capabilities that are present in source, included in the Desktop composition, and reachable through a user workflow. Longer-term ideas are kept separate.

Near-term roadmap

These directions do not yet have complete first-class product journeys and are not counted as current features.

Project overview

DeepSeek Harness Studio uses Electron to host the DeepSeek Harness Web workspace. The desktop main process starts and manages a local dsh web service. This repository provides the complete development source so users can clone or download it, install dependencies, edit the code, launch the desktop app, and continue development.

Desktop installers are published only through this repository's GitHub Releases page, never through a third-party download site. Electron-validated macOS arm64 and Windows x64 development previews are available now, while the complete source remains available for local development.

Workspace and Agent execution

• Workspaces and sessions: choose a local directory natively; group, search, or unregister Workspaces; rename, archive, or fork a session at its last completed turn.
• Planning and work management: organize work with Plan, Goal, and Todo. The Jobs panel shows background work owned by the current process; a process restart does not present those old jobs as still running.
• Workflows and SubAgents: chat reconstructs Workflow phases, members, and outcomes. The SubAgent directory exposes parent/child lineage, continuable conversations, and stopping the current turn of a running continuable child.
• References and deliverables: @file and @session supply context. Successfully produced files appear at the end of the answer and can be opened or revealed through the local Host.
• Human collaboration: Agents can ask structured single-choice, multiple-choice, or custom questions. Tool approvals, full access, and plan review require an explicit UI decision.
• Security boundaries: permission presets bind sandbox and approval policy to each session. Credentials are stored through a write-only flow; the page never reads back or displays a stored secret value.

DeepSeek Harness v0.1.1-rc.2 compatibility

Studio 0.1.0-rc.19 integrates the core and Web capabilities from DeepSeek Harness 0.1.1-rc.2 while retaining Beyondata's Plugin Center, Plugin Discovery, Preset Square, Application Center, themes, and desktop recovery flows. Studio and upstream Harness versions are managed independently; the download links above match this release.

• Multimodal capability: retains Pro and Flash text models while adding DeepSeek-V4-Flash-Vision-Exp, persistent image attachments, and reusable Files API uploads; stale references retry within bounds and file-resolution failure falls back for the whole request to bounded inline images.
• Agent runtime: integrates @ file/session references, Plan, Goal, background Jobs, Workflow, SubAgents, concurrent Web Search, and persistent PowerShell PTY sessions on Windows.
• Desktop adaptation: starts the Host with --no-open, retains native directory selection, plugin transaction recovery, and existing user-data locations, and performs lockfile-free compatibility recovery for incompatible historical plugin locks.

Plugin ecosystem: discover what is worth installing, then manage it

Plugin Discovery: start here when you do not know what to install

Not sure where to find plugins, which ones were updated recently, or what the ecosystem is paying attention to? Open Plugin Discovery from the sidebar. The app reads the online catalog automatically and turns scattered packages into a recommendation page you can browse and act on directly.

• A fresh place to start every day: see catalog features, recently updated entries, and ecosystem-popular plugins without searching repositories one by one.
• Filter by scenario: browse Agents and workflows, Web UI, browser and search, vision and media, memory and context, models and services, developer tools, or integrations and notifications.
• Search for the answer directly: search by plugin name, capability keyword, or publisher, then inspect its icon, summary, version, and update time.
• Act as soon as you discover it: start the trusted installation flow for a new plugin, or jump to Plugin Center management for an installed one.

Do not know the exact package name? Let the Agent shortlist it

When all you know is “I want a desktop pet,” you do not need to guess an npm package name first. Enter the need in Plugin Discovery and the app sends it to the current Agent as a /find-plugins request. The Agent loads the built-in Skill, performs a read-only search of the public dsh-plugin catalog, and returns the closest candidates with versions, publishers, update dates, and matching reasons in the current conversation.

• No catalog vocabulary required: describe the outcome, use case, or problem in ordinary language.
• Evidence stays inspectable: each result includes the exact package name, version, publisher, update date, and a matching reason.
• Search and installation remain separate decisions: recommendations are public-catalog metadata only; after choosing a package, use Plugin Center for compatibility checks and installation confirmation.

Plugin Center: install, enable, disable, and remove online

After choosing a plugin, open Plugin Center and search by short package name, full npm name, or an explicit GitHub repository for plugins and Skill Packs published to the public npm Registry in the DeepSeek Harness Bundle format. The dsh-plugin keyword is only a discovery signal, and GitHub is only used to map back to a published npm package; source is never installed directly.

• Online discovery: search public plugins and inspect versions, capabilities, permissions, compatibility, and risk.
• One-click installation: download and verify an exact package version, integrity metadata, and Bundle declaration; after confirmation, Desktop installs it, restarts the Harness Host, and verifies runtime state.
• Installed management: review system, public-catalog, and local sources together, then enable, disable, update, or uninstall an entry from its three-dot menu.
• Safe removal: uninstall retains configuration and plugin data by default; deleting data requires a separate user confirmation.

Built-in skins and custom backgrounds

Open Settings → Background to switch built-in skins. For a custom image, the app performs the 1920×1080 WebP crop and interface color adaptation locally without uploading the original.

Models, permissions, and thinking modes

• Permission selection: the composer uses the Chinese 只读, 工作区写入, and 完全访问 labels for the current session. General settings affect only new sessions, and enabling Full access requires an explicit risk confirmation.
• Model and thinking modes: the model and API key remain managed in Settings. The composer shows the current DeepSeek model and offers 关闭思考, 低强度思考, 深度思考, and 最大思考.

Local models and self-hosted inference

Studio can connect to model services running on the same machine or local network through standard OpenAI-compatible interfaces. Users start the inference server and load the model first; Studio does not download model weights, allocate deployment storage, or manage GPU runtime parameters.

• Conversation models: open Settings → Models → Add local model, choose Ollama, vLLM, SGLang, or OpenAI-compatible, then fetch or enter a model id; the API base is prefilled and the local API key is optional.
• Local vision models: choose Ollama, vLLM, SGLang, or a custom service in vision-enhancement setup; the configuration is saved only after a real image succeeds.
• No silent cloud fallback: a failed self-hosted route never forwards the image to Bailian, OpenRouter, or another cloud provider.

Vision enhancement: let DeepSeek understand images

When vision enhancement is enabled, the Host selects one route from the current exact model's capabilities: a model that declares image support receives the original image; otherwise, a configured and verified Bailian or OpenRouter-compatible provider supplies a traceable observation to the Agent. The routes are mutually exclusive, so each image is processed once.

• Available in the composer: use the “视觉增强” shortcut on the left side of the input bar; hover to see its purpose and current state.
• Automatic path selection: the shortcut reports 原生 or 兼容 · 提供方; native vision needs no extra key, while compatible vision requires a verified provider.
• Explicitly off: disabling the switch keeps images out of model-visible context while retaining attachment history in the interface.
• Built for development work: understand product screenshots, error dialogs, designs, charts, photos, and text in images, or inspect an image by its workspace path.

Download the desktop app

GitHub Releases provides Electron-validated macOS Apple Silicon and Windows x64 development previews. Running either desktop build requires no separate Node.js or pnpm installation. These remain development-preview assets; formal releases will provide platform-signed macOS .dmg and Windows x64 .exe installers.

Development previews use a separate pre-release tag without triggering the formal installer workflow. Their public download area contains only the macOS ZIP and Windows installer; checksums, blockmaps, and platform-verification records remain in the corresponding GitHub Actions artifact so users do not mistake development files for installers. The formal workflow accepts only a desktop-v* tag that exactly matches the Desktop version, and publishes the macOS and Windows installers with SHA256SUMS only after both platform signatures pass verification.

Quick start

Get the source

Clone the repository with Git:

git clone https://github.com/fufankeji/deepseek-harness-studio.git
cd deepseek-harness-studio

You can also choose Code → Download ZIP on the GitHub repository page, extract the archive, and open the project directory.

Requirements

• Node.js ^22.19.0 || >=24.0.0
• pnpm 11.7.0

External services

Downloading the source, installing dependencies, and launching the desktop development environment do not require an API key. Configure the selected model provider and credentials in the application settings only when making model requests, and never commit credentials to Git.

Install and run

Install the workspace dependencies:

pnpm install

Build the required modules and launch the desktop development environment:

pnpm run dev:desktop

The development launcher rebuilds when relevant source or build inputs change. To force a complete rebuild, run:

pnpm run dev:desktop:rebuild

Repository layout

deepseek-harness-studio/
├── apps/
│   ├── desktop/       # Electron main process, preload, Host lifecycle, and desktop build scripts
│   ├── web/           # DeepSeek Harness Web entry and desktop composition
│   └── cli/           # dsh CLI, runtime configuration, and Agent Presets
├── packages/          # Agent, model, tool, session, plugin, and client capability packages
├── native/            # Native sandbox helpers
├── python/            # Python SDK and related runtime
├── examples/          # Runnable examples and configurations
├── scripts/           # Build, validation, generation, and publishing scripts
├── website/           # Documentation site source
├── vendor/            # Pinned Cordis foundation source
└── assets/            # Project images used by the README

Common development commands

Suggested reading order

1. apps/desktop/src/main.ts: desktop entry, window, tray, and local Host composition.
2. apps/desktop/src/host-supervisor.ts: dsh web startup, readiness detection, and shutdown.
3. apps/desktop/src/preload.ts: fixed desktop interfaces exposed to the renderer.
4. apps/web/: the Web workspace loaded by the desktop window.
5. apps/cli/ and packages/: CLI composition and Harness capabilities.

Relationship to DeepSeek Harness

This project continues desktop development from the Harness core, Cordis plugin system, and Web interface in deepseek-ai/deepseek-harness. This repository maintains the Electron desktop entry, local Host management, desktop interactions, and supporting development scripts.

License

This project uses the MIT License. Third-party license information is available in THIRD_PARTY_NOTICES.md.