
Turn any OpenAI-compatible model into a workspace-aware desktop agent — Skills, tools, sub-agents, and a local, auditable request log.
Rabbit Chat Studio is a Tauri 2 + Rust + React desktop client. It is not tied to a model vendor: any endpoint that speaks the OpenAI Chat Completions API can be used. Beyond chat, it gives the agent a real workspace: skills, files, commands, memory, todos, timers, MCP servers and sub-agents all operate inside one directory boundary, and every request and tool call is recorded locally.
These pages document Rabbit Chat Studio v3.3.3. Every English page has a Simplified Chinese counterpart; tool names, file names and config keys stay in English in both languages.
Typical use cases:
- multi-model switching and prompt engineering;
- contextual conversations with searchable history;
- semi-automated tasks that combine local commands, files and web scraping;
- MCP integration and tool-ecosystem expansion.
5-minute quick start
- Install and launch: on Windows, the Microsoft Store listing installs and updates it for you; otherwise download the installer for your platform from Latest release. On macOS, right-click the app → Open the first time.
- Configure a model: open Settings → API & Model, enter the base URL and API key, click Fetch Models From API, pick a model and save.
- Choose a workspace: open Settings → Workspace & Profiles and select your project directory. Rabbit Chat Studio creates a hidden
.data/folder there for chats, skills, memory and todos. - Start chatting: type a message and press Enter; Shift+Enter inserts a newline. Enable tools from the Tools panel when you want the model to act on local files or run commands.
See Quick Start for the full first-run walkthrough.
Feature map
| What you want to do | Page |
|---|---|
| Install, configure an API, first chat | Quick Start |
| Build from source, package, publish | Installation and Build |
Chat, model parameters, @file / /skill mentions, reasoning, context, drawings, PDF export |
Interface and Chat |
| New/search/favorite/archive/fork sessions | Sessions and History |
Switch workspaces, understand .data and app data |
Workspace and Data |
| Write a skill, restrict its tools and commands | Skills |
| Enable tools, understand risk levels and confirmations | Tools Overview |
| Read/write files, run commands and workspace scripts, background serve | Files and Commands |
| Run slow tools, MCP calls or sub-agents in the background | Files and Commands, Tools Overview |
| Memory, todos, delayed timers | Memory, Todos and Timers |
| Add MCP servers | MCP Servers |
| Configure LLM / CLI / cloud sub-agents, mission bus and planner takeover | Sub-Agents |
| System prompt, context compression, reasoning | Prompts and Context |
config.json, profiles, export/import |
Configuration and Backup |
| Request monitor, tool logs, mission monitor | Monitoring and Logs |
| In-app updates and manual upgrades | Updates |
| Path boundaries, permissions, key safety | Security and Boundaries |
| Code structure, commands, CI, roadmap | Development and Architecture |
| Something is broken | FAQ |
Why Rabbit Chat Studio
Chat is the easy part. Rabbit Chat Studio focuses on three things that are easy to miss: giving the agent a real workspace, giving it usable local tools, and keeping an auditable record of every call — without imposing a persona.
| Capability | Rabbit Chat Studio | Cherry Studio | Chatbox | LM Studio |
|---|---|---|---|---|
Skill sandbox — one skill root per workspace (.data/skills), root pinned to its SKILL.md |
✅ | ◐ agent workspaces | — | — |
| Self-evolution — the assistant installs/edits/deletes its own skills and keeps verified facts in a dedicated memory scope | ✅ | — | — | — |
Built-in agent utilities — 4-level memory (session / user / repo / evolution), timers, todos |
✅ | ◐ scheduled tasks | — | — |
| Agent runtime — LLM, local CLI or remote cloud, configured in-app | ✅ | ◐ LLM agents | — | — |
| Prompt ownership — no default persona; your system message is the base, tool/skill guidance is appended | ✅ | ◐ assistant prompts | ◐ prompt library | ◐ per-model prompt |
| Local request audit — body, tokens, latency, status, errors | ✅ | ◐ history only | ◐ history only | ◐ history + server log |
| Background execution — slow tools, MCP calls and sub-agents keep running while the model works | ✅ | — | — | — |
| Built-in drawing — Mermaid, flow charts, 2D SVG and 3D scenes, saved as PNG/HTML | ✅ | — | — | — |
✅ built in · ◐ partial or adjacent · — not offered. Based on public documentation as of November 2026.
README coverage
| README section | Where it lives here |
|---|---|
| Why Rabbit Chat Studio / Introduction | This page |
| Core Features → Chat | Interface and Chat, Sessions and History |
| Core Features → Sub-Agent Orchestration | Sub-Agents |
| Core Features → Skill System | Skills |
| Core Features → Tool System | Tools Overview, Files and Commands, Memory, Todos and Timers |
| Core Features → MCP Management | MCP Servers |
| Core Features → Local Persistence & Operations | Workspace and Data, Configuration and Backup, Monitoring and Logs |
| First Principles Design | Development and Architecture |
| Tech Stack & Architecture | Development and Architecture |
| Quick Start | Quick Start |
| Build & Installation | Installation and Build |
| Configuration | Configuration and Backup |
| Project Structure | Development and Architecture, Workspace and Data |
| Security & Boundaries | Security and Boundaries |
| Development Roadmap | Development and Architecture |
| License | Development and Architecture |
Design principles
The core intent: reason from first principles, return freedom to the user (tools, LLM API, prompts), and keep the interface minimal and professional.
Rabbit Chat Studio reduces a desktop agent to four irreducible problems: how to reliably obtain model output, how to let the model act on the local world, how to draw clear risk boundaries, and how to keep everything reproducible and debuggable.
- Minimal dependency closure: model calls go through OpenAI-compatible APIs, so there is no vendor lock-in.
- Capabilities decoupled from permissions: chat works by default; tools are opt-in; skills can narrow them further.
- Local-first observability: configuration, history and request logs stay local.
- Explicit security boundaries: skill directory isolation, tool allowlists and workspace path constraints.
- Replaceable architecture: frontend, command layer, model service and graph backend are modular.
In one sentence: the model handles cognition and planning, local tools handle execution and verification, and the application framework handles boundaries and records.
Conventions
- Settings → API & Model refers to the sections in the Settings panel.
- Paths are relative to the current workspace root unless stated otherwise;
./is the workspace root. - Commands show Windows PowerShell first; macOS/Linux differences are called out.
- Tool names, config keys and event names stay in English so they are easy to search in
config.jsonand the source.
Keeping these pages current
These docs live in the repository under docs/wiki/ and are rendered to this site by scripts/gen-wiki.mjs — this site is the only maintained copy. The GitHub Wiki now holds forwarding pages that point here. Version numbers in the pages follow package.json (v3.3.3 today) and the per-release entries in CHANGELOG.md.
Links
- Repository: https://github.com/xlbbb-cn/ai-chat
- Microsoft Store: https://apps.microsoft.com/detail/9MV82J9G95GS (Store ID
9MV82J9G95GS) - Latest release: https://github.com/xlbbb-cn/ai-chat/releases/latest
- Changelog: https://github.com/xlbbb-cn/ai-chat/blob/main/CHANGELOG.md
- Issues: https://github.com/xlbbb-cn/ai-chat/issues
- License: MIT (see
LICENSEin the repository)