# Pinvou Agent

> Open-source desktop agent workspace for work, design and coding

- Category: Agents & Automation / Assistants & Chatbots
- Language: Rust
- Licence: MIT
- Repository: https://github.com/Pinvou/pinvou-agent
- Website: https://pinvou.com
- Canonical page: https://ai-tldr.dev/tools/pinvou-agent/

## In short

Pinvou Agent is a desktop application for Windows, macOS and Linux that treats a chat as a task meant to end in a result: a file, a visual or a code change. It sits in the Assistants & Chatbots category of the Agents & Automation tools and is open-source, written in Rust and licensed MIT.

## Overview

Pinvou Agent is a desktop application for Windows, macOS and Linux that treats a chat as a task meant to end in a result: a file, a visual or a code change. It splits that work into three modes. Work combines attachments, a personal knowledge base, specialist personas, Skills, MCP tools and workflows to research, analyse and write reusable files. Design turns a prompt into a poster or data visualisation that you can then edit directly — selecting elements and changing copy, fonts, colours and layout — or keep refining by description. Code brings Codex, Claude Code or Kimi into a real project (or an isolated temporary workspace) through the Agent Client Protocol (ACP), showing plans, tool steps, permission requests and file changes.

Under the hood it is a Tauri 2 app with a React and Vite interface. The agent engine is CodeWhale, a separate submodule that handles model calls, streaming, tool execution, sessions, MCP, Skills, hooks and context compaction; the desktop layer owns the UI, configuration and operating-system integration. Every file the agent creates or edits is collected in an artifact panel for preview, and Markdown artifacts can be edited in place or revised by selecting a passage. Plan and YOLO modes choose between reviewing a plan first and executing directly.

Pinvou Agent works with a local vLLM server or any OpenAI-compatible API, with built-in templates for DeepSeek, Kimi, Qwen, Doubao, MiniMax, Zhipu GLM, MiMo, OpenAI, Anthropic, Gemini and xAI. Sessions, settings, knowledge and extensions live under `~/.pinvou3/`; whether data leaves the machine depends on which model, MCP servers and connectors you enable. The project is MIT-licensed and published preview builds through GitHub Releases.

## What it does

- Three modes in one workspace: Work for file deliverables, Design for editable posters and data visualisations, Code for project work with a coding agent
- Codex, Claude Code or Kimi driven over ACP, with plans, tool steps, permission requests and file changes shown in the app and sessions bound to their workspace
- Artifact panel that collects every file the agent creates or edits, with in-place editing and passage-level revision for Markdown
- Local knowledge base with full-text and vector retrieval, multiple collections per chat and source provenance in answers
- Memory center with explicit review of captured preferences, plus a reusable persona card pool
- Unified tool store for local and remote MCP servers, CLI tools and API connectors, with OAuth/SSO where supported and ready-made connectors such as Feishu, DingTalk, WeCom and Obsidian
- Attachments for PDF, Office documents, images and text; local voice input; GPU, memory and context-usage monitoring
- Plan and YOLO modes, and an auxiliary side chat that can only read the running task's history
- Local vLLM or any OpenAI-compatible endpoint, with per-session switching between saved model configurations

## Getting started

Preview installers are published on the GitHub Releases page: a universal .dmg for macOS 11 or later, a Windows x64 setup .exe, and .deb packages for Linux x64, arm64 and LoongArch64 (Ubuntu 22.04 or newer). The steps below follow the project README.

1. Install and connect a model

   Download the installer for your platform from GitHub Releases and launch it. In the application settings, add a model configuration — a built-in template such as DeepSeek, Qwen, OpenAI, Anthropic or Gemini, a local vLLM server, or any custom OpenAI-compatible endpoint.

2. Or point it at a local vLLM server

   Endpoints can also be supplied through environment variables. For a non-loopback plain-HTTP endpoint on a trusted network, also set DEEPSEEK_ALLOW_INSECURE_HTTP=1.

```bash
export DEEPSEEK_BASE_URL="http://127.0.0.1:8000/v1"
export DEEPSEEK_API_KEY="local-no-auth"
export DEEPSEEK_MODEL="your-model-name"
```

3. Pick a mode and give it a task

   Choose Work, Design or Code below the input box. In Code mode, pick Codex, Claude Code or Kimi and select a project directory, or type directly to start a temporary session. Files the agent produces appear in the artifact panel.

4. Run from source

   Requires Git with submodule support, Node.js and npm, a current Rust toolchain and the Tauri 2 system dependencies for your platform.

```bash
git clone --recursive https://github.com/Pinvou/pinvou-agent.git
cd pinvou-agent/pinvou3-app
npm ci
cd ..
./pinvou3-app/run-dev.sh
```

## When to use it

- Reach for it when a task should end in a file — a report, meeting summary or spreadsheet — rather than a chat reply
- Reach for it to drive Codex, Claude Code or Kimi from one desktop app with visible plans, permissions and file changes
- Reach for it to keep a fully local loop with a vLLM server, local knowledge base and local tools
- Reach for it to wire Chinese workplace tools such as Feishu, DingTalk, WeCom or Tencent Docs into an agent through the tool store
