Pinvou Is an Open-Source Desktop AI Agent for Work, Design, and Code

Most AI agent tools live in a browser tab or a chat window, and the moment you want them to touch your actual files, run a coding agent, or persist a knowledge base across sessions, you are back to gluing separate tools together yourself. Pinvou is an open-source desktop AI agent built to close that gap directly. Rather than another chat interface, it is a workspace app that unifies research, visual design, and coding agents into one environment that lives on your own machine, connects to your own tools over MCP, and produces actual files instead of chat replies you have to copy out somewhere else.

The project is new, created July 24, 2026, and moving fast: over 700 GitHub stars and nearly 100 forks in under three weeks, with a commit landing as recently as this morning. It is MIT licensed and free.

The three modes

Pinvou organizes around three distinct modes rather than one generic chat window, and the split is the actual design decision worth understanding before you install it.

Work mode is the closest to a conventional agent: you combine file attachments, a local knowledge base, saved personas, and tools to research and analyze something, and the output is meant to be a reusable file, not just an answer in a chat log. Design mode turns a natural-language prompt into an editable visual artifact, a poster or a data visualization you can then adjust directly rather than re-prompting from scratch. Code mode is where Pinvou differs most from a typical desktop AI wrapper: it integrates real coding agents, including Codex, Claude Code, and Kimi, through something called the ACP protocol, and lets the agent work inside your actual project directories rather than a sandboxed toy environment.

How it is actually built

Under the hood, Pinvou is a layered system rather than a single script. A React and Vite frontend talks to a desktop orchestration layer over Tauri, and that layer hands off to CodeWhale, a separate agent engine submodule that handles model calls, tool execution, MCP integration, and session management. The split matters practically: UI and system-integration contributions go into the app layer, while engine improvements follow their own fork policy, which is a reasonable structure for a project this young trying to keep its core agent logic separable from its desktop shell.

MCP is the connective layer that makes the workspace idea real. Pinvou ships a unified tool store integrating MCP servers, CLI tools, and API connectors, so a tool you already use with another MCP-compatible agent can plug straight into any of the three modes rather than needing a Pinvou-specific integration. It also includes a memory center for retaining long-term context and preferences across sessions, local voice input with downloadable speech models, and OAuth and SSO support for enterprise chat tools like Feishu, DingTalk, and WeCom, which signals it is being built with more than a solo hobbyist in mind despite its age.

Model access and what running it actually looks like

Pinvou is not tied to one model provider. It supports local vLLM deployments and any OpenAI-compatible API endpoint, with built-in templates for DeepSeek, Kimi, Qwen, Doubao, and other providers, and you can configure multiple model profiles and switch between them per session.

Getting it running today is a developer-grade setup, not a one-click install. It requires Git with submodule support, Node.js and npm, the Rust toolchain, and Tauri's system dependencies, then:

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

It runs on Windows, macOS 11 and later on Apple Silicon, and Linux. All settings, sessions, and knowledge bases are stored locally in ~/.pinvou3/.

Privacy and honest limitations

Whether Pinvou is actually private depends entirely on what you configure it to talk to. Run it with local models and local tools, and it stays fully private on your machine. Connect it to cloud models or remote connectors, and that data leaves your machine the same way it would with any other cloud-backed agent. The project does not paper over this distinction, and it is worth taking seriously before pointing it at anything sensitive.

The project's youth shows in a few places worth knowing before you commit time to it. There is no packaged installer, so setup means cloning a repo with submodules and building it with a real development toolchain, which puts it out of reach for anyone who is not comfortable in a terminal. With over 40 open issues against a three-week-old repository, expect rough edges. And because Code mode leans on separate coding agents like Claude Code and Codex rather than reimplementing that capability itself, you still need those tools set up and working on their own for that mode to be useful.

Who Pinvou is actually for

Pinvou is not the right pick if you want a polished, install-and-go assistant, or if you are not comfortable building from source. It is a strong pick if you already run several separate AI tools, an MCP server here, a coding agent there, a chat app somewhere else, and want one workspace that treats those as first-class citizens instead of forcing everything through a single chat box. For anyone building a personal operator setup around AI, the combination of MCP-native tool access, a real coding-agent bridge, and local-model support is a genuinely different shape than most desktop AI apps on offer right now, even if the install process makes clear this is still an early build.

Sources: Pinvou Agent, GitHub