darknoah/aime-chat

---

✨ Features

• 🦾 Harness Engineering - A complete agent harness around the model (Agent = Model + Harness): orchestration loop, tools, context & memory, sub-agents, guardrails, and observability turn a stateless model into a reliable, long-running agent
• 🤖 Multiple AI Provider Support - Integrated with mainstream AI providers including OpenAI, DeepSeek, Google, Zhipu AI, MiniMax, Ollama, LMStudio, ModelScope, and more
• 💬 Intelligent Conversations - Powerful AI Agent system based on Mastra framework, supporting streaming responses and tool calling
• ⏯️ Message Replay - Review saved conversation parts, including compressed history, with adjustable playback speed, pause/resume, and progress seeking; assistant replies show elapsed time when timestamps are available
• 🤝 Open CoWork Capability - AI is not just for chatting: it can edit project files in an integrated workspace, execute code, search the web, and more
• 📚 Knowledge Base Management - Local knowledge bases with hybrid semantic/BM25 retrieval, optional reranking, source-text inspection, document/Excel parsing, and long-term cultivation memory
• 📤 Project Chat Export - Select project threads and export complete histories as Markdown, raw JSON, Excel, or Unsloth JSONL
• 🧠 Cultivation Memory - A scheduled Cultivation Agent extracts preferences, habits, project context, and important facts from chat history into a structured memory wiki
• ⏰ Cron Automation - Run scheduled AI tasks with project context, selectable agents/tools, and either reusable or per-run chat threads
• 🛠️ Tool Integration - Support for MCP (Model Context Protocol) client with extensible tool capabilities
• 🎙️ Audio Processing - Local and provider-backed speech recognition and synthesis, including Alibaba ASR/TTS and MiniMax ASR; MiniMax Music generates songs or instrumental tracks and saves the audio locally
• 🎬 Video Generation - Generate videos with MiniMax H3 / H3 Max or Alibaba Wan 3.0, using text, frames, or reference media supported by the selected model
• 🧊 3D Generation & Preview - Generate GLB models from text or images with Alibaba Tripo, then inspect GLB, glTF, FBX, OBJ, or STL files in the project workspace with material and wireframe views
• 🔍 Skill System - Search, import, and manage AI skills from Git repositories or the online skill marketplace
• 🧑‍💻 Assistant Personalities - Built-in assistant personalities can be selected instantly and customized through the current personality format
• 🖥️ Background Bash Sessions - Track long-running shell processes from the current chat or the whole project, with direct stop controls in the UI
• 🌐 Embedded Browser - Electron Chromium with shared sign-in data, multiple tabs per chat, and isolated actions across concurrent chats
• 📡 Channel Integration - Connect AI capabilities to messaging platforms like WeChat and Telegram
• 🔐 Secrets Management - Centrally manage credentials for tools and services and inject them into configured local workflows
• 🎨 Customizable UI - Choose light, dark, or system mode, set an accent color, and configure independent sidebar/chat backgrounds
• 🌍 Internationalization - Built-in Chinese and English interfaces
• 🔒 Local First - App state and chat data are stored locally by default; cloud providers still receive the requests you send to them
• ⚡ High Performance - Built on Electron for cross-platform native experience

🚀 Quick Start

Prerequisites

• Node.js >= 22.x
• npm >= 10.x
• pnpm >= 10.x

Install Dependencies

pnpm install

Development Mode

Start the development server:

• Click Electron: Main in VSCode’s debug panel (or Electron: All to attach both processes)

The application will start in development mode with hot reload support.

Build Application

Package desktop application:

pnpm package

Packaged applications will be generated in the release/build directory.

macOS Installation Notes

Due to the app not being signed with an Apple Developer certificate, macOS Gatekeeper may prevent the app from running. If you see “App is damaged” or “Cannot be opened” error, please run the following command in Terminal:

# After mounting the DMG and copying to Applications
xattr -cr /Applications/aime-chat.app

Or right-click the app → hold Option key → click “Open”.

📦 Project Structure

aime-chat/
├── assets/              # Static assets
│   ├── icon.png        # Application icon
│   ├── models.json     # AI model configurations
│   └── model-logos/    # Provider logos
├── src/
│   ├── main/           # Electron main process
│   │   ├── providers/  # AI provider implementations
│   │   ├── mastra/     # Mastra Agent and tools
│   │   ├── knowledge-base/ # Knowledge base management
│   │   ├── tools/      # Tool system
│   │   └── db/         # Database
│   ├── renderer/       # React renderer process
│   │   ├── components/ # UI components
│   │   ├── pages/      # Page components
│   │   ├── hooks/      # React Hooks
│   │   └── styles/     # Style files
│   ├── types/          # TypeScript type definitions
│   ├── entities/       # Data entities
│   └── i18n/           # Internationalization config
└── release/            # Build artifacts

🎯 Core Features

Harness Engineering

A raw LLM is just a stateless function — it becomes a dependable agent only when wrapped in a harness. Following the Agent = Model + Harness formula that the industry formalized in 2026, AIME Chat is built as a complete harness around the model, not just a chat box in front of it.

The harness layers AIME Chat provides:

In short, AIME Chat focuses on engineering everything around the model so that any provider — cloud or local — can be turned into a reliable, goal-directed agent.

AI Provider Configuration

Support for configuring multiple AI providers, each with independent settings:

• API Key
• API Endpoint
• Available model list
• Enable/Disable status

Supported providers include:

Knowledge Base Features

• 📄 Import text, Markdown, PDF, Word, Excel, PowerPoint, and OCR-readable images
• 🔍 Hybrid semantic and BM25 full-text retrieval, with BM25-only operation when no embedding model is configured
• 🎯 Optional reranking and safe full-index rebuilding when changing the embedding model
• 📖 Inspect the original source after retrieval; long sources support pattern search and paginated reading
• 💡 Intelligent Q&A based on knowledge base
• 🧠 Cultivation memory that maintains a global memory wiki from chat history
• 📊 Knowledge base management interface

Cultivation Memory

AIME Chat includes a global memory knowledge base maintained by the built-in Cultivation Agent. When the Cultivation Daily cron task is enabled, it reads newly updated user conversations, filters out automation-generated threads, deduplicates against existing memories, and writes useful long-term information into Markdown pages such as preferences.md, habits.md, and project notes.

This helps future conversations automatically inherit stable preferences, working habits, important people/entities, and ongoing project context without pasting old chat logs into every prompt.

Message Replay

After the current response finishes, click Replay messages in the chat toolbar and choose an interval of 0.5, 1, 3, or 5 seconds. Replay reveals saved text, reasoning, tool, and file parts in sequence, merging available compressed history with the current conversation and removing duplicate message IDs. Use the controls to pause/resume, adjust speed, seek, replay from the beginning after completion, or exit.

Replay displays stored messages without rerunning the Agent or its tools. Switching chats or starting a new response exits replay automatically. Reply elapsed time is calculated from the preceding user message and the assistant’s last-part timestamp; it is omitted when timestamps are missing or invalid.

See the basic usage guide for replay and compressed-history controls.

Tool System

Rich built-in tools that AI Agents can call autonomously:

• 🧵 Background Bash Tracking - Long-running Bash sessions are visible from task/context surfaces, including project-wide sessions when a chat is bound to a project
• 🔌 MCP Protocol Support - Extensible third-party tools
• ⚙️ Tool Configuration UI - Visual tool management and configuration
• 🔍 Skill Marketplace - Search and import skills from Git repositories or online marketplace (skills.sh)

Media Tool Setup

Configure and enable the provider first. Alibaba media tools use Workspace ID and Region from provider settings; chat continues to use its configured API Base. Tripo currently requires cn-beijing. Choose video, transcription, and speech models in Settings → Default Models, with per-tool overrides when needed. Select 3D and music models in their respective tool configurations; existing music-2.5 configurations need a supported replacement.

Generation tools wait for the result and save it to the workspace. Stopping ends local waiting; provider-side cancellation varies by service. See the tool guide for supported inputs and the provider guide for setup.

Channel Integration

Connect AI capabilities to external messaging platforms:

Secrets Management

Centralized management of secret keys and credentials used by tools and services:

• 🔑 Unified interface for managing API keys and tokens
• 💾 Values are stored in the local application database; protect your OS account, database, and backups
• 🔗 Automatic injection into tools that require authentication

🛠️ Tech Stack

Frontend

• Framework: React 19 + TypeScript
• UI Library: shadcn/ui (based on Radix UI)
• Styling: Tailwind CSS
• Routing: React Router
• State Management: React Context + Hooks
• Internationalization: i18next
• Markdown: react-markdown + remark-gfm
• Code Highlighting: shiki

Backend (Main Process)

• Runtime: Electron
• AI Framework: Mastra
• Database: TypeORM + better-sqlite3
• Vector & Full-text Storage: LibSQL
• Embeddings: Configurable local or provider-backed embedding models
• AI SDK: Vercel AI SDK

Build Tools

• Bundler: Webpack 5
• Compiler: TypeScript + ts-loader
• Hot Reload: webpack-dev-server
• App Packaging: electron-builder

Project Initialization

git clone https://github.com/DarkNoah/aime-chat.git
cd ./aime-chat
pnpm install

# Since pnpm disables postinstall scripts by default, if you encounter missing binary packages or similar issues, run:
pnpm approve-builds

⚙️ Configuration

Optional Runtime Libraries

AIME Chat supports optional runtime libraries that can be installed from the Settings page:

These runtimes are installed under the application data directory. Runtime install attempts write detailed success and failure information to the application log, and the About page provides a direct entry for opening the log file.

The Bash tool can use either AIME Chat’s independent Python runtime (recommended) or the system Python runtime. Configure this per tool when an Agent needs packages or environment behavior from a specific Python installation.

Appearance and Browser Instances

• Open Settings → Appearance to choose the theme mode and accent color, or add separate JPG/PNG/WebP backgrounds for the sidebar and chat area. Background opacity and blur can be adjusted independently.
• Open Settings → Instances to view the shared Electron Chromium browser, its data directory and open tab counts, or close all browser tabs. Browser data lives in /instances/default_browser; no separate browser installation is required.

Data Storage

Application data is stored by default in the system user directory:

• macOS: ~/Library/Application Support/aime-chat
• Windows: %APPDATA%/aime-chat
• Linux: ~/.config/aime-chat

🤝 Contributing

Issues and Pull Requests are welcome!

1. Fork this repository
2. Create a feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

Code Standards

• Use ESLint and Prettier to maintain consistent code style
• Follow TypeScript type specifications

📄 License

This project is licensed under the MIT License.

👨‍💻 Author

Noah

• Email: [email protected]

🙏 Acknowledgments

• Electron
• React
• Mastra
• Vercel AI SDK
• shadcn/ui
• Radix UI

🔗 Related Links

• Official Website
• Issue Tracker
• Releases and Changelog

---