DeepDepCat

Dual-workflow AI desktop workbench — one native application for both coding and document/office automation, natively optimized for DeepSeek V4.

version

Why DeepDepCat

There are terminal coding agents (they only write code) and chat clients (they only talk). DeepDepCat is neither — it is both, in a single native desktop app:

Code workspace — LSP-driven coding agent: subagent orchestration, multi-session parallel conversations, execution strategies (plan-execute / reflexion / coordinator / generate-review), checkpoint rollback, sandboxed tool execution, and permission modes from "confirm every change" to "fully automatic".
Depwork workspace — document & office automation agent: Word/PPT/Excel generation, OCR, table processing, media conversion, desktop UI automation, batch file operations — with a live task panel showing step-by-step progress, pausable and resumable at any time.

DeepSeek V4 is not bolted on — it is the core: 1M context window, max reasoning effort, streaming thinking-mode display, and context-cache hit metrics you can actually see.

This repository is the open-source home of the DeepDepCat desktop app, released under Apache-2.0. It is also the release channel and community hub. Issues, feature requests and feedback are very welcome.

Screenshots

Onboarding Code workspace Depwork workspace

Quick start

Download the latest Windows installer from Releases, or visit the official site.
Open the app — Settings → Model Providers, select the built-in DeepSeek provider, and paste your DeepSeek API Key.
Add models deepseek-v4-pro / deepseek-v4-flash (1M context window by default).
Start working — Code workspace for code, Depwork workspace for documents.

Key features

Feature	Description
Dual workflows	Coding + document/office automation in ONE app, each with its own tool chain, system prompt and agent definitions
Multi-session parallelism	Sessions run concurrently without blocking each other; live per-session indicators and one-click stop in the sidebar
Per-session terminal	A persistent interactive shell per chat session — switch sessions to switch terminals; supports python/vim/htop
Subagent orchestration	Complex tasks decomposed into parallel subagents with real-time activity tracking
Execution strategies	Standard / plan-execute / reflexion / coordinator / generate-review (independent evaluator)
Pause & resume	Tasks pause at checkpoints (context fully preserved) and resume anytime — not a kill
DeepSeek V4 native	1M context, max reasoning effort, streaming thinking display, cache-hit metrics
Long-term memory	Cross-session project memory with automatic injection and Chinese-semantic retrieval (FTS5 CJK)
Skills & ecosystem	Own directory-based skills plus Claude/Cursor skill & plugin compatibility — zero format conversion
MCP support	First-class MCP (stdio / SSE / HTTP); external server tools merge into one registry
Checkpoint rollback	Undo any agent file change with one click — back to the pre-edit state
Trust & safety	3-level permission modes (read-only / accept edits / allow all), sandboxed execution, bash AST security analysis, inline user confirmations
Silent updates	Small releases install automatically on exit, zero UI; major releases via the update button

Building from source

DeepDepCat is fully open source (Apache-2.0). To build the desktop app yourself:

Install Rust, Node.js 18+, and the Tauri 2 prerequisites.
Run npm install.
Run npm run tauri build (or npm run tauri dev for development).

Source layout: src — React frontend · src-tauri — Rust backend · depwork-mcp — WPS Office MCP server.

WPS Office MCP server

WPS Office MCP server (depwork-mcp/) — control WPS Office (Writer / Calc / Impress) from any MCP-compatible AI client: document creation & editing, export to docx/xlsx/pptx/pdf, and live typewriter-style writing into a visible window. JSON project model + deferred COM rendering — editing never requires WPS to run.

Add it to your MCP client in one block:

{
  "mcpServers": {
    "wps-office": {
      "command": "python",
      "args": ["-m", "wps_controller.mcp_server"]
    }
  }
}

See depwork-mcp/README.md for the full tool list, CLI usage and JSON project format.

Platform support

Windows (x64) — available now (NSIS installer, silent updates)
macOS (Apple Silicon / Intel) — cross-platform Tauri app; builds in progress
Linux — planned

Tech stack

Desktop shell: Tauri 2 (Rust backend + WebView2 / WebKit)
Frontend: React 18 + TypeScript + Vite + Zustand
Cloud backend (not in this repo): Python FastAPI + SQLite (updates, auth, telemetry, sync)
Models: DeepSeek V4 (default), OpenAI, Anthropic, xAI Grok, local Ollama — OpenAI / Anthropic / Responses protocols