flowllm 0.2.0.3

FlowLLM: Simplifying LLM-based HTTP/MCP Service Development

FlowLLM: Simplifying LLM-based HTTP/MCP Service Development
_{If you find it useful, please give us a ⭐ Star. Your support drives our continuous improvement.}

English | 简体中文

---

📖 Introduction

FlowLLM encapsulates LLM, Embedding, and vector_store capabilities as HTTP/MCP services. It is suitable for AI assistants, RAG applications, and workflow services, and can be integrated into MCP-compatible client tools.

🏗️ Architecture Overview

🌟 Applications Based on FlowLLM

Project Name	Description
ReMe	Memory management toolkit for agents

📢 Recent Updates

Date	Update Content
2025-11-15	Added File Tool Op feature with 13 file operation tools, supporting file reading, writing, editing, searching, directory operations, system command execution, and task management
2025-11-14	Added Token counting capability, supporting accurate calculation of token counts for messages and tools via self.token_count() method, with support for multiple backends (base, openai, hf). See configuration examples in default.yaml

📚 Learning Resources

Project developers will share their latest learning materials here.

Date	Title	Description
2025-11-14	HaluMem Analysis	HaluMem: Evaluating Hallucinations in Memory Systems of Agents Analysis
2025-11-13	Gemini CLI Context Management Mechanism	Multi-layer Context Management Strategy for Gemini CLI
2025-11-10	Context Management Guide	Context Management Guide
2025-11-10	LangChain&Manus Video Materials	LangChain & Manus Context Management Video

⭐ Core Features

• Simple Op Development: Inherit from BaseOp or BaseAsyncOp and implement your business logic. FlowLLM provides lazy-initialized LLM, Embedding models, and vector stores accessible via self.llm, self.embedding_model, and self.vector_store. It also offers prompt template management through prompt_format() and get_prompt() methods. Additionally, FlowLLM includes built-in token counting capabilities. Use self.token_count() to accurately calculate token counts for messages and tools, supporting multiple backends (base, openai, hf, etc.).

• Flexible Flow Orchestration: Compose Ops into Flows via YAML configuration. >> denotes serial composition; | denotes parallel composition. For example, SearchOp() >> (AnalyzeOp() | TranslateOp()) >> FormatOp() builds complex workflows. Define input/output schemas and start the service with flowllm config=your_config.

• Automatic Service Generation: FlowLLM automatically generates HTTP, MCP, and CMD services. The HTTP service provides RESTful APIs with synchronous JSON and HTTP Stream responses. The MCP service registers as Model Context Protocol tools for MCP-compatible clients. The CMD service executes a single Op in command-line mode for quick testing and debugging.

---

⚡ Quick Start

📦 Step0 Installation

📥 From PyPI

pip install flowllm

🔧 From Source

git clone https://github.com/flowllm-ai/flowllm.git
cd flowllm
pip install -e .

For detailed installation and configuration, refer to the Installation Guide.

⚙️ Configuration

Create a .env file and configure your API keys. Copy from example.env and modify:

cp example.env .env

Configure your API keys in the .env file:

FLOW_LLM_API_KEY=sk-xxxx
FLOW_LLM_BASE_URL=https://xxxx/v1
FLOW_EMBEDDING_API_KEY=sk-xxxx
FLOW_EMBEDDING_BASE_URL=https://xxxx/v1

For detailed configuration, refer to the Configuration Guide.

🛠️ Step1 Build Op

from flowllm.core.context import C
from flowllm.core.op import BaseAsyncOp
from flowllm.core.schema import Message
from flowllm.core.enumeration import Role

@C.register_op()
class SimpleChatOp(BaseAsyncOp):
    async def async_execute(self):
        query = self.context.get("query", "")
        messages = [Message(role=Role.USER, content=query)]

        # Use token_count method to calculate token count
        token_num = self.token_count(messages)
        print(f"Input tokens: {token_num}")

        response = await self.llm.achat(messages=messages)
        self.context.response.answer = response.content.strip()

For details, refer to the Simple Op Guide, LLM Op Guide, and Advanced Op Guide (including Embedding, VectorStore, and concurrent execution).

📝 Step2 Configure Config

The following example demonstrates building an MCP (Model Context Protocol) service. Create a configuration file my_mcp_config.yaml:

backend: mcp

mcp:
  transport: sse
  host: "0.0.0.0"
  port: 8001

flow:
  demo_mcp_flow:
    flow_content: MockSearchOp()
    description: "Search results for a given query."
    input_schema:
      query:
        type: string
        description: "User query"
        required: true

llm:
  default:
    backend: openai_compatible
    model_name: qwen3-30b-a3b-instruct-2507
    params:
      temperature: 0.6
    token_count: # Optional, configure token counting backend
      model_name: Qwen/Qwen3-30B-A3B-Instruct-2507
      backend: hf  # Supports base, openai, hf, etc.
      params:
        use_mirror: true

🚀 Step3 Start MCP Service

flowllm \
  config=my_mcp_config \
  backend=mcp \  # Optional, overrides config
  mcp.transport=sse \  # Optional, overrides config
  mcp.port=8001 \  # Optional, overrides config
  llm.default.model_name=qwen3-30b-a3b-thinking-2507  # Optional, overrides config

After the service starts, refer to the Client Guide to use the service and obtain the tool_call required by the model.

---

📚 Detailed Documentation

🚀 Getting Started

• Installation Guide
• Configuration Guide

🔧 Op Development

• Op Introduction
• Simple Op Guide
• LLM Op Guide
• Advanced Op Guide
• Tool Op Guide
• File Tool Op Guide
• Vector Store Guide

🔀 Flow Orchestration

• Flow Guide

🌐 Service Usage

• HTTP Service Guide
• HTTP Stream Guide
• MCP Service Guide
• CMD Service Guide
• Client Guide

---

🤝 Contributing

Contributions of all forms are welcome! For participation methods, refer to the Contribution Guide.

---

📄 License

This project is licensed under the Apache 2.0 license.

---

Star 历史

---

GitHub • Documentation • PyPI