案板 Anban

什么都能下锅的创作台。
The creation bench where everything goes into the pot.

English · 简体中文

案板 (Anban) is a local-first content creation workbench for food / history / natural-history creators, built on top of the DeepSeek Harness desktop shell. Put your raw material on the board — cutting, mincing, stir-frying, plating are Anban's job.

It combines three things:

• DSH Desktop shell — a fork of dataelement/dsh-desktop (MIT), a cross-platform desktop wrapper around DeepSeek Harness.
• Free search stack — dsh-search-free: multi-layer web search (Exa → Tavily → Bing auto-fallback) + web_fetch, no paid official search package needed.
• Creator skill pack — a five-step creation pipeline distilled from real creator workflows.

Important

Anban is an early preview pinned to @deepseek-ai/dsh@0.1.0-rc.6. Current builds are not code-signed or notarized by Apple and are not recommended for production use.

The creation pipeline

Say “帮我写个口播稿” (help me write a voice-over script) and Anban walks the full pipeline:

Step	Skill	What it does
1	topic-mining	Mine topics from your material pool & trends into topic cards (hook / audience / angle / series / verifiability)
2	fact-check	Multi-source verification of every concrete fact (numbers, dates, names) via web_search, with confidence levels
3	oral-script	Write the script in a 老饕 (foodie) + storyteller tone: hook → story arc → spoken-language rewrite → golden line
4	storyboard	Turn the script into a shot list (visuals / captions / SFX / pacing) with a footage checklist
5	publish-checklist	Pre-publish & post-publish checklist for Bilibili + YouTube (titles, covers, descriptions, series, 48h review)

Skills live in skills/ and are auto-loaded by the anban-creator agent preset (老饕 + 说书人 persona, grounded, never fabricates history).

Download

Platform	Package	Download
macOS Apple Silicon	DMG installer	Download for Apple Silicon
macOS Intel	DMG installer	Download for Intel Mac
Windows x64	Setup installer	Download Windows installer
Windows x64	Portable executable	Download portable version

All current and historical packages are on the GitHub Releases page.

Quick start (new install)

1. Install the DMG, right-click → Open (unsigned build).
2. Choose a workspace.
3. The default model is jy (tokenrhythm relay) — configure per profile-overrides/settings.model.yml (keys via JY_API_KEY env, never committed).
4. web_search works out of the box with the free search stack — configure per profile-overrides/cordis.patch.yml and symlink the plugin into the parent tree (never npm install inside a live profile).
5. Start a session on the 案板创作助手 preset and say “帮我写个口播稿”.

See profile-overrides/README.md for both environments (web profile ~/.dsh/profiles/web/, desktop profile under ~/Library/Application Support/dsh-desktop/harness/profiles/).

Development

git clone https://github.com/lmcsh9527/anban.git
cd anban
npm install        # runs patch-package + brand-asset install + Electron runtime
npm run dev        # electron-vite dev

Quality checks:

npm test
npm run typecheck
npm run build
node scripts/check-runtime-deps.mjs   # 19 runtime @deepseek-ai deps present

Packaging (run on the matching platform/arch):

npm run package:mac:arm64
npm run package:mac:x64
npm run package:win
scripts/smoke-packaged.sh <path-to-Anban.app>   # verify the 19 runtime deps shipped

CI (.github/workflows/release.yml) builds macOS arm64 + Intel + Windows x64, checks the 19 runtime deps, and publishes a GitHub Release on tags.

Runtime architecture

The upstream shell architecture is unchanged: Anban is a fork of dataelement/dsh-desktop, which launches a local Harness instance on a random 127.0.0.1 port, persists profiles/plugins/sessions outside the app install directory, and hardens the BrowserWindow. See the upstream README for the full architecture.

Upstream collaboration

• This repository is a fork of dataelement/dsh-desktop (MIT). The upstream MIT attribution is kept (see LICENSE).
• Upstream bug fixes are contributed back as PRs — e.g. PR #10 (declare the 19 runtime @deepseek-ai deps so electron-builder bundles them).
• The main branch stays clean for upstream PRs; Anban development lives on anban-brand.

Related projects

Project	Description
dsh-search-free	Free multi-layer search + fetch plugin (Exa → Tavily → Bing + web_fetch), published on npm
dataelement/dsh-desktop	Upstream desktop shell (MIT)

License

MIT © lmcsh9527, on top of the dataelement/dsh-desktop MIT base (upstream attribution retained).

DeepSeek Harness and its dependencies remain subject to their respective upstream licenses and trademark policies. Anban is an independent community project and is not affiliated with DeepSeek.