Sharpclaw

中文版

A .NET 10 AI agent with long-term memory and system operation tools. Supports Anthropic Claude, OpenAI, and Gemini multi-provider switching, with Alibaba Cloud DashScope for vector embedding and reranking.

Features

• Multi-provider support — Anthropic / OpenAI / Gemini, switch via config without code changes
• Per-agent configuration — Each agent (main, memory recall, memory save, summarizer) can use different API endpoints, keys, and models
• Long-term memory — Automatically saves, retrieves, and injects important information from conversations, persisted across sessions
• Memory pipeline — Four independent sub-agents collaborate: proactive memory saving, memory recall injection, sliding window trimming, conversation summarization
• Vector semantic search — Two-phase retrieval: vector embedding recall + optional reranking
• Semantic deduplication — Automatically merges memories when cosine similarity exceeds threshold, avoiding redundancy
• System tools — File operations, process execution (dotnet/node/docker), HTTP requests, background task management
• TUI interface — Terminal.Gui v2 based interface with separate chat, log, and input areas

Requirements

• .NET 10 SDK

Quick Start

dotnet run --project sharpclaw

First run automatically launches the configuration wizard. Select your AI provider, enter API keys, and configure per-agent settings.

Config is saved to ~/.sharpclaw/config.json. After that, launching goes straight to interactive chat. Type /exit or /quit to exit, Ctrl+Q to quit.

Configuration

Run the config wizard:

dotnet run --project sharpclaw config

Config file structure (~/.sharpclaw/config.json):

{
  "version": 4,
  "default": {
    "provider": "anthropic",
    "endpoint": "https://api.anthropic.com",
    "apiKey": "sk-xxx",
    "model": "claude-opus-4-6"
  },
  "agents": {
    "main": {},
    "recaller": { "enabled": true },
    "saver": { "enabled": true, "model": "claude-haiku-4-5-20251001" },
    "summarizer": { "enabled": true, "model": "claude-haiku-4-5-20251001" }
  },
  "memory": {
    "enabled": true,
    "embeddingEndpoint": "https://dashscope.aliyuncs.com/compatible-mode/v1",
    "embeddingApiKey": "sk-xxx",
    "embeddingModel": "text-embedding-v4",
    "rerankEnabled": true,
    "rerankEndpoint": "https://dashscope.aliyuncs.com/compatible-api/v1/reranks",
    "rerankApiKey": "sk-xxx",
    "rerankModel": "qwen3-vl-rerank"
  }
}

Each agent inherits from default unless overridden. Set "enabled": false to disable a sub-agent.

Memory Pipeline

Processing flow per conversation turn:

User Input
  |
  +- MemoryRecaller    Recall: retrieve relevant memories, inject into context
  |
  +- Agent Response    Main agent processes (can call SearchMemory/GetRecentMemories tools)
  |
  +- SlidingWindowChatReducer (AfterMessageAdded)
       +- MemorySaver            Proactive save: analyze conversation, save/update/delete memories
       +- Strip old injected messages
       +- Sliding window trim    Triggered when exceeding windowSize + buffer
       +- ConversationSummarizer Summarize trimmed conversation, inject summary

Data Persistence

• ~/.sharpclaw/config.json — Provider and agent configuration
• history.json — Session state, auto-restored on startup
• memories.json — Vector memory store