<div align="center">

askr

MCP Q&A Assistant / MCP 问答助手

Async Q&A capabilities for AI clients via MCP protocol, powered by any OpenAI-compatible API.

通过 MCP 协议为 AI 客户端提供异步问答能力，后端对接任意 OpenAI 兼容 API。

English | 中文文档

🤖 100% Vibe Coding — Built entirely through AI-assisted development.

⭐ If you like this project, please give it a star!

</div>

---

Quick Start

# First time: launch the management panel to configure your provider
npx @sweatent/askr -m

# Start the MCP stdio server
npx @sweatent/askr

Features

• question — Ask a single question with streaming support
• agentq — Ask multiple independent questions in parallel (configurable concurrency)
• list — View recent session statuses
• check — Retrieve session results with blocking wait, full content, and chain-of-thought support

MCP Client Integration

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "askr": {
      "command": "npx",
      "args": ["askr"]
    }
  }
}

Claude Code

Add via CLI or global config:

claude mcp add askr -- npx @sweatent/askr

Or manually edit .claude/settings.json:

{
  "mcpServers": {
    "askr": {
      "command": "npx",
      "args": ["askr"]
    }
  }
}

Cursor / Other MCP Clients

Configure a stdio-type MCP server with command npx @sweatent/askr per your client's documentation.

Management Panel

npx @sweatent/askr -m
# or
npx @sweatent/askr --manage

Menu options:

1. Manage Provider — Configure Base URL, API Key, model (auto-fetch model list or manual input)
2. Manage Questions — View/close active sessions
3. View Logs — Browse session history with question, chain-of-thought, and answer (live refresh for running sessions)
4. More Settings — Max concurrency, timeout, fold characters, system prompt, language

Supports Chinese and English interfaces.

Configuration

Config file is stored in the system data directory:

Platform	Path
Windows	%AppData%/askr/config.json
macOS	~/Library/Application Support/askr/config.json
Linux	~/.config/askr/config.json

{
  "language": "en",
  "provider": {
    "baseUrl": "https://api.example.com",
    "apiKey": "sk-xxx",
    "model": "gpt-4",
    "stream": true
  },
  "settings": {
    "maxConcurrent": 5,
    "timeout": 150,
    "foldChars": 1000,
    "systemPrompt": "You are a search assistant..."
  }
}

Base URL

• Standard: https://api.example.com — automatically appends /v1/chat/completions
• Custom endpoint: https://api.example.com/custom/path# — trailing # means use as-is (strip the #)

Tool Reference

question

{ "content": "What is REST and its core principles?" }

Ask a single question. Answers exceeding foldChars are truncated; use check for full content. Returns session ID on timeout.

agentq

{ "questions": ["What is Docker?", "What is Kubernetes?"] }

Parallel questions, up to maxConcurrent. Each question runs in isolated context, returning its own result or session ID.

list

{ "count": 5 }

Returns the most recent N session summaries (ID, timestamp, status, question preview).

check

{ "id": "a3f8k2", "showFull": true, "showThinking": true }

Retrieve session result. Blocks if the session is still running. Parameters:

• showFull — Return full answer without truncation
• showThinking — Include chain-of-thought content

Highlights

• Streaming — SSE streaming with real-time session file updates
• Auto fallback — Automatically falls back to non-streaming if streaming fails
• Chain-of-thought separation — Detects <think> tags, reasoning_content, and thinking fields
• Hot reload — Config changes from the management panel take effect immediately, no restart needed
• Session persistence — All Q&A records saved locally with full history browsing

Requirements

• Node.js >= 18

Recommended Prompt

Add the following to your project's CLAUDE.md or other AI client system prompt for best results:

## askr Usage Guidelines

You can use askr MCP tools to search and ask an external AI to assist with your tasks.

### When to Use

- **Pre-implementation research**: Before using unfamiliar or uncertain public APIs / library functions, query their signatures, parameters, usage, and caveats to avoid writing incorrect code from memory
- **Unfamiliar concepts**: Look up terms, protocols, or design patterns you don't fully understand
- **Latest information**: Query recently released versions, changelogs, news, or announcements that may not be covered in your training data
- **Troubleshooting**: When facing hard-to-solve errors or unexpected behavior, search for solutions including GitHub Issues, Stack Overflow discussions, etc.
- **Solution comparison**: When deciding between multiple technical approaches, query their trade-offs and community recommendations
- **Configuration & compatibility**: Look up platform-specific configurations or known compatibility issues for particular environments or versions

### Principles

1. Prefer using askr to search first; only rely on your own knowledge when you are fully confident it is accurate and up-to-date
2. Be specific in your questions, include necessary context (language, framework, version, etc.), avoid vague queries
3. Combine related sub-questions into one question; use agentq for unrelated parallel questions
4. Use check to retrieve timed-out results; re-ask if check times out twice
5. Use check(id, showFull: true) when you need the full untruncated answer

Acknowledgements

Thanks to the LinuxDo community for the support!

License

MIT