LangClaw

OpenCLaw-equivalent AI agent platform built on LangGraph. Phase 1: Core agent runtime with ReAct loop, model abstraction, tool registry, skills, and MCP.

Phase 1 Features

• LangGraph Agent: StateGraph with ReAct loop (agent → tools → agent)
• Model Abstraction: OpenAI, Anthropic, Google Gemini, Ollama (config-driven)
• Built-in Tools: calculator, get_current_time, web_search, read_file, write_file, shell (sandboxed), send_email, calendar_add, calendar_list
• Skills: Load SKILL.md from skills/<name>/ folders; inject instructions into system prompt
• MCP Integration: Load tools from MCP servers (e.g. firecrawl-mcp) via config
• Checkpointer: MemorySaver for conversation persistence
• YAML Config: Agent definition in config.yaml

Quick Start

1. Install (from de-LLM root, or install langclaw):

   cd applications/langclaw
   pip install -e .

2. Run Ollama (for local model):

   ollama serve
   ollama pull llama3.2:3b

3. Run agent:

   python run.py
   # Or: python run.py default

Telegram Channel

1. Install telegram extra:

cd applications/langclaw
pip install -e ".[telegram]"

2. Create a Telegram bot with BotFather and set:

export TELEGRAM_BOT_TOKEN="..."

3. Run the Telegram adapter:

source ../../.venv/bin/activate
python run_telegram.py default

Each Telegram chat_id maps to a stable LangClaw thread_id (telegram:<chat_id>), so memory persists per chat in langclaw.sqlite.

Phase 2: Memory (SQLite)

LangClaw now persists short-term conversation history to applications/langclaw/langclaw.sqlite keyed by thread_id.

• Resume a previous session: set LANGCLAW_THREAD_ID before running:

export LANGCLAW_THREAD_ID="my-session-1"
python run.py

• Demo of “history in DB as new input”:

python demo_db_history.py

This script does:

• load previous messages from SQLite (load_messages(thread_id))
• append them to the new user message
• invoke the agent
• write the new turn back to SQLite (append_messages(thread_id, ...))

Config

Edit config.yaml:

agents:
  default:
    model: ollama/llama3.2:3b
    system_prompt: "You are a helpful assistant."
    tools:
      [
        calculator,
        get_current_time,
        read_file,
        write_file,
        shell,
        send_email,
        calendar_add,
        calendar_list,
      ]
    skills: [example]
    temperature: 0.7
    # Optional MCP tools (requires: pip install mcp langchain-mcp-adapters)
    # mcp:
    #   server: npx
    #   args: [firecrawl-mcp]
    #   env:
    #     FIRECRAWL_API_KEY: ${FIRECRAWL_API_KEY}

Model format: provider/model (e.g. openai/gpt-4o, anthropic/claude-sonnet, ollama/llama3).

Skills

Add skills in skills/<name>/SKILL.md. Each SKILL.md is injected into the system prompt. Enable via skills: [name1, name2] in config.

MCP

Install optional deps: pip install langclaw[mcp]. Configure MCP server in agent config to load additional tools (e.g. Firecrawl for web scraping).

Project Structure

langclaw/
├── config.yaml
├── run.py
├── pyproject.toml
└── src/langclaw/
    ├── config/       # YAML loader
    ├── models/       # Model providers (OpenAI, Anthropic, Ollama, Google)
    ├── runtime/      # LangGraph agent graph
    └── tools/        # Registry + builtin tools

Knowledge about　langgraph

the MCP is a tool the RAG is a tool retriever as a node, it will add the data from embedding db to system prompt

vector size and how to measure the similarity of content.

telegram service is saparate process, you just use the LLM invoke to response message.

Test

uv run pytest