A Desktop AI Agent Built to Deliver Files, Not Just Chat

What it does

Pinvou Agent is a Tauri-based desktop workspace that wraps a local or remote LLM and surrounds it with tools for producing actual work. You can assign tasks in a work mode with attachments and knowledge bases, generate editable posters and data visualizations in a design mode, or embed external coding agents such as Codex and Claude Code into real projects through an ACP connector. Everything persists in multi-session workspaces with an artifact panel that collects files the agent creates, rather than letting them vanish into chat history.

The interesting bit

The project treats chat as a means, not an end. Its most unusual move is the code mode, which does not try to be the coder itself; instead, it hosts established coding agents inside the same desktop shell via ACP, giving them access to project files and terminal commands while keeping the session bound to a workspace. That is a rare admission that specialized agents are already good at coding, so the app acts as an orchestration layer rather than a replacement.

Key highlights

• Three distinct modes—Work, Design, and Code—each with tailored tooling and artifact handling.
• Local-first option: run a local vLLM instance with local tools and keep data on the machine.
• Unified tool store for MCP servers, CLI connectors, and API integrations, including ready-made connectors for Feishu, DingTalk, Obsidian, and Tencent services.
• Built-in knowledge base with full-text and vector retrieval, plus a memory center for long-term preferences.
• Remote control via QR code to steer the workspace from a phone.

Caveats

• The README trails off mid-sentence, which is a fair signal that the project is still heavily in flux.
• It does not bundle a language model; you must provide your own OpenAI-compatible endpoint or local vLLM.
• The agent engine lives in a separate CodeWhale submodule, so the codebase is split across repositories and governed by its own fork policy.

Verdict

Worth a look if you want a desktop control tower for multiple AI modes and care about local-first operation or Chinese enterprise integrations. Skip it if you need a polished, batteries-included product or a simple chat wrapper.