Skip to content

README.md

Latest commit

 

History

History
92 lines (70 loc) · 4.85 KB

README.md

File metadata and controls

92 lines (70 loc) · 4.85 KB

Remote MCP Client Quickstart (Hosted Only)

This guide is for users who only consume the hosted MCP server. You do not need to clone or navigate the full langgraph-dev-navigator repo. Follow these steps in your own workspace.

What You Need

  • Hosted MCP URL: {deployed-remote-http-mcp-server-url} (ask the maintainer privately; the real URL is not in this repo).
  • MCP client config for your AI assistant (Claude, GitHub Copilot, etc.).
  • LangGraph rules file: langgraph-ai-rules_v4_1.md (recommended for remote).
  • Optional: local copies of LangChain/LangGraph/LangSmith docs in your own project.

Step 1: Get the Hosted MCP URL

  • Contact the maintainer to obtain the actual URL. Keep it private.

Step 2: Configure Your AI Assistant (MCP Server)

Claude CLI

claude mcp add --transport http crawl4ai-rag-http-remote {deployed-remote-http-mcp-server-url}

GitHub Copilot

.vscode/mcp.json

{
  "servers": {
    "crawl4ai-rag-http-remote": {
      "url": "{deployed-remote-http-mcp-server-url}",
      "type": "http"
    }
  },
  "inputs": []
}
  • Copilot note: ASK mode cannot access MCP; use PLAN mode (or equivalent) when you need MCP tools.
  • Connection tip: Copilot PLAN mode may take several minutes (sometimes ~7 mins) to connect to the hosted MCP server, especially because the server is a serverless deployment that may cold-start. Prompt the AI to run a simple MCP tool (e.g., get_available_sources) to confirm the connection once it comes up.

Other MCP-capable editors (Cursor/Windsurf/etc.)

  • Add a server entry pointing to {deployed-remote-http-mcp-server-url} with type/http (or your editor’s MCP schema).

Step 3: Apply the LangGraph Rules (langgraph-ai-rules_v4_1.md)

Choose one:

  • Option A (script; write into a workspace you choose):

    uv run python src/setup_dev_assistant.py --rule-file langgraph-ai-rules_v4_1.md --workspace /path/to/your/project

    Then pick your assistant from the menu. The script writes the assistant config files under the given workspace (e.g., /path/to/your/project/.vscode/mcp.json, AGENTS.md, etc.).

  • Option B (recommended for your own workspace):

    • Download the file directly (see langgraph-ai-rules_v4_1.md locally or on GitHub) or copy its content into a local file.
    • Use your assistant’s UI to attach the file in the chat/session.

Delivery advice (important):

  • Many assistants ignore system rule files dropped into default paths. The reliable method—and the one that actually empowers the assistant to follow the LangChain-specific workflow—is to attach the rule file inside the chat so it becomes part of the conversation context. Example for Copilot PLAN mode:
    “Give me a plan to implement a web search multi-agent with checkpoint feature. I attach langgraph-ai-rules_v4_1.md for how to work with LangChain ecosystem tasks.”
  • Reminder: Copilot ASK mode cannot access MCP; use PLAN mode (or equivalent) when you need MCP tools.
  • Want to see realistic workflows? Review these step-by-step examples that show how users finish LangGraph tasks with the hosted MCP server:

Step 4 (Recommended): Local Docs in Your Workspace

If you want local copies of the docs (not required for hosted MCP use), you can add them to your project:

git clone https://github.com/langchain-ai/langgraph.git langgraph
git clone https://github.com/langchain-ai/docs.git docs
git clone https://github.com/langchain-ai/langsmith-docs.git langsmith-docs

This matches the doc map referenced by the LangGraph rules.

You’re Ready

  • Use your assistant with MCP configured to the hosted server and the attached langgraph-ai-rules_v4_1.md.
  • No other folders from langgraph-dev-navigator are needed for remote-only usage.

Notes & Troubleshooting

  • check_ai_script_hallucinations currently has a known bug for remote MCP environments and is disabled; rely on other MCP tools until a fix is deployed.
  • Because the MCP server is hosted on a serverless platform, cold starts can increase connection time. If a tool request stalls, wait and rerun (starting with a light-weight call such as get_available_sources).
  • For best local search performance when you clone docs yourself, install ripgrep (sudo apt install ripgrep on Debian/Ubuntu or brew install ripgrep on macOS).