RemindAI is an open-source desktop AI assistant built around a complete ToolShell layer that gives LLMs the ability to manipulate files, execute code, call external tools, manage persistent memory, and autonomously plan tasks — turning AI into a productivity tool that can actually do things, not just talk about them.
🎯 Beyond the chatbox — give AI real agency.
| 🔵 Typical AI Client | 🟣 RemindAI | |
|---|---|---|
| 📁 File Ops | ❌ Not supported | ✅ Built-in sandboxed filesystem |
| 💻 Code Exec | ❌ Not supported | ✅ Built-in Python/Shell/JS executor |
| 🧠 Memory | ❌ None or context-only | ✅ Vector semantic memory + SQLite + soft-failure filter |
| 🔌 Extensions | ✅ MCP + four-layer Skills + Capability plugins | |
| 🤝 Multi-Agent | ✅ Real collaboration with routing & permission isolation + parallel doc comprehension | |
| 🔄 AgentLoop | ❌ None | ✅ Controllable cyclic pipeline: Think→Write→Test→Verify |
| 📦 Skill Import | ✅ One-click ZIP import + batch import | |
| 🌐 External API | ❌ Not supported | ✅ Built-in HTTP API server with three endpoint types |
| 🐱 Desktop Companion | ❌ None | ✅ Pixel pet + TTS voice + shop economy + achievements |
RemindAI's Skill System uses a four-layer architecture, each with independent storage and lifecycle:
| Layer | Name | Storage | Lifecycle | Description |
|---|---|---|---|---|
| L1 | Default Meta-Skills | assets/default_skills/ |
Global, shipped with app | ToolShell, Schedule, System — the three core meta-skills forming AI's fundamental capabilities: file I/O, command execution, task planning, environment probing |
| L2 | User Global Skills | Skills/ |
Global, user-toggled | Imported via ZIP or created with /skill-cti; reusable across projects. Format: SKILL.md + tools.json |
| L3 | Workspace Temp Skills | .toolshell/skills/ |
Per workspace, always active | AI creates on-demand during guidance; solidifies workflows for the current project; disappears when switching directories — never pollutes global skills |
| L4 | AI Self-Generated Skills 🧪 | (Planned) | Global, not yet implemented for safety | AI auto-generates skills from long-term conversation memory (e.g., if you frequently consult on operations research, AI distills a dedicated OR skill) and invokes it autonomously |
- L1 Meta-Skills: The AI's "OS kernel" — file I/O, command execution, environment probing, task scheduling; the foundation of ToolShell
- L2 Global Skills: Your "toolbox" — reusable expertise for specific domains, code generation, document templates, workflow automation
- L3 Temp Skills: The AI's "sticky notes" — solidify a workflow for the current project, discard cleanly when done. For example, the ToolShell/Schedule/System meta-skill definitions in
memory.jsonare injected via the L3 mechanism - L4 Self-Generated (planned): The AI's "long-term learning" — distill domain preferences and working patterns from conversations into personalized skills. Deferred due to safety concerns around auto-generated executable code
| Command | Purpose | Destination |
|---|---|---|
| Direct request to create a skill | Create a project-level skill in current workspace | L3 .toolshell/skills/ (default) |
/skill-temp |
Explicitly create a project-level temp skill | L3 .toolshell/skills/ |
/skill-cti |
Create → Self-test → Install as global skill | Built in .toolshell/_staging/, installed to L2 Skills/ after passing tests |
💡 If RemindAI's skills system inspires your projects, papers, or other research, please help me improve and link to the project. This would be very helpful for my graduation and future employment. 🙇
| Module | Status | Notes |
|---|---|---|
| AI Chat Core (LLM + tool calling) | ✅ | AgentLoop streaming cycle + event-driven UI |
| Controllable AgentLoop Pipeline | ✅ | Think → Write → Test → Verify cyclic pipeline |
| Three LLM Protocols (OpenAI/Anthropic/Gemini) | ✅ | Independent clients, streaming+tools+multimodal |
| ToolShell Meta-Skill | ✅ | read/write/delete/search/exec/python/js + rg/fd/rtk |
| Schedule Meta-Skill | ✅ | 7 tools CRUD + review + archive |
| System Meta-Skill | ✅ | Env probe + sanitized env vars |
| MCP Multi-Transport | ✅ | stdio / SSE / Streamable HTTP |
| Vector Memory | ✅ | Qdrant + SQLite dual-write + auto failover + soft-failure filtering |
| Pluggable Capability | ✅ | Search landed, framework extensible |
| Four-Layer Skill System | ✅ | L1 default meta + L2 user global + L3 workspace temp + L4 planned, batch import support |
| Model Card Management | ✅ | CRUD + logo + drag-sort |
| Multi-Agent Collaboration | ✅ | Framework complete + parallel doc comprehension orchestration + controllable AgentLoop pipeline |
| Domain Experts | ✅ | Preset/custom roles + skill binding |
| Conversation Export | ✅ | MD / PDF / Word / HTML |
| Desktop Experience | ✅ | Tray / notifications / splash / theme animation |
| Global Pet Agent | ✅ | Pixel cat + TTS voice + shop economy + achievements |
| External API Server | ✅ | Built-in HTTP server, three endpoints: OpenAI aggregation / Claude Agent / Claude proxy |
| Online Agent Access | ✅ | Remote access to RemindAI Agent via browser |
| Context Compression | ✅ | RTK Token compression 60-90% + context management optimization |
| Flowchart Summary | ✅ | Use archify to summarize conversations as flowcharts |
| Feature | Description |
|---|---|
| 🐚 ToolShell | File sandbox + Python/Shell/JS exec + rg/fd/rtk + RTK compression 60-90% token savings |
| 🌐 API Server | Built-in HTTP server with three endpoints: OpenAI aggregation, Claude Agent (runs RemindAI's own agent loop), and Claude proxy (pass-through) |
| 🔌 MCP Protocol | stdio/SSE/Streamable HTTP + auto-discovery + drag-and-drop management |
| 🧠 Vector Memory | Qdrant semantic search + SQLite backup + auto-ops + soft-failure filtering + index rebuild |
| 🤝 Multi-Agent | Commander/Worker/Reviewer roles + permission isolation + auto-routing + parallel doc comprehension |
| 🔄 Controllable AgentLoop | Think → Write → Test → Verify cyclic pipeline with long-conversation stutter prevention |
| 🎨 Multi-Model | OpenAI/Anthropic/Gemini native + streaming reasoning chain + multimodal |
| 🧩 Capability | Pluggable architecture, Custom → MCP → ToolShell three-tier routing |
| 📦 Skills | Four-layer architecture (L1 meta / L2 global / L3 temp / L4 self-gen planned), SKILL.md + tools.json format, one-click ZIP import + batch import, command-based creation |
| 🔍 Web Search | Tavily / Brave / Baidu AI Search, session-level toggle |
| 📋 Schedule | SCHEDULE.md driven, P0/P1/P2 priority, AI proactive review |
| 👤 Domain Experts | Preset/custom roles + dedicated system prompts |
| 🖼️ Built-in Tools | Gemini image gen / Formula OCR / PaddleOCR / Flowchart / Rich-text |
| 📊 Flowchart | Use archify to summarize conversations as flowcharts |
| 📤 Export | Markdown / PDF / Word / HTML |
| 🌍 i18n | Full Chinese and English |
| 🎨 Themes | Material 3 light/dark + ripple transition animation |
| 🐱 Global Pet Agent | Pixel cat companion + right-click AI Q&A + Volcano TTS + shop/inventory/feeding + achievements |
| 🗜️ Context Compression | RTK output compression + intelligent conversation context trimming |
| 🌐 Online Access | Remote browser access to Agent with online session management |
The app ships with these executables — no extra installation needed:
| Tool | Description | Source |
|---|---|---|
rg |
ripgrep — blazing fast regex search | BurntSushi/ripgrep |
fd |
fd — modern file finder | sharkdp/fd |
rtk |
RTK — Token compressor, 60-90% output reduction | nicobailey/rtk |
Head to Releases for pre-built packages:
| Platform | Status | Notes |
|---|---|---|
| 💻 Windows | ✅ Officially supported | Installer available |
| 🐧 Linux | 🔧 Build from source | Compiles and runs fine |
| 🍎 macOS | 🔧 Build from source | Compiles and runs fine |
# Requirements: Flutter SDK >= 3.12.1
git clone https://github.com/PythonnotJava/RemindAI.git
cd RemindAI
# Windows
flutter build windows --release --tree-shake-icons --split-debug-info=./debug-info
# Linux
flutter build linux --release --tree-shake-icons --split-debug-info=./debug-info
# macOS
flutter create --platforms=macos
flutter build macos --release --tree-shake-icons --split-debug-info=./debug-info📸 Click to expand
| Feature | Screenshot |
|---|---|
| 🏠 Main Interface |  |
| 📁 Working Directory |  |
| 🔌 MCP Services |  |
| 🧠 Memory System |  |
| 🤝 Multi-Agent |  |
| 📦 Skills System |  |
Thanks to Yu for designing the delightful logo that brings life and personality to RemindAI.
- https://arxiv.org/pdf/2606.24775 — Thanks to this paper for pinpointing a known weakness in memory architectures: the lack of version management leads to retrieval of stale facts.
- Is it possible to design a tool paradigm like this: tool name, brief description, version, and documentation URL (so the model can look up unfamiliar tool commands on the fly), allowing the Agent to auto-inject them when relevant?
If RemindAI helps you, feel free to support development ~
💚 WeChat 🔵 Alipay
MIT License — Copyright (c) 2026 PythonnotJava
Built with Flutter and passion ❤️