mcpgram-python-sdk

Official Python SDK for MCPGRAM — one client for native connectors and external MCP servers, with LangGraph and CrewAI adapters.

Mirrors @mcpgram/sdk (the JS/TS SDK) field-for-field against the same /api/v1/tools and /api/v1/execute endpoints, so both SDKs stay interchangeable.

Install

pip install mcpgram              # core client only
pip install "mcpgram[langgraph]" # + native LangChain/LangGraph tool objects
pip install "mcpgram[crewai]"    # + native CrewAI tool objects
pip install "mcpgram[all]"       # everything

Quickstart

from mcpgram import Platform

client = Platform(api_key="mcpg_live_...", base_url="https://your-app.vercel.app")

github = client.use("github")
result = github.call("github_list_repos", {"per_page": 10})
print(result.output)

LangGraph

toolset = client.use("github")
lg = toolset.for_langgraph()

lg.tools          # plain {name, description, parameters} dicts — zero extra deps
lg.langchain_tools # actual langchain_core.tools.StructuredTool instances
                    # (only populated if langchain-core is installed; otherwise None)

Drop lg.langchain_tools straight into a LangGraph ToolNode or agent's tool list — each one calls back into MCPGRAM's /api/v1/execute under the hood.

CrewAI

toolset = client.use("notion")
cw = toolset.for_crewai()

cw.tools        # plain {name, description, args_schema} dicts
cw.crewai_tools # actual crewai.tools.BaseTool instances (needs `crewai` installed)

Design notes

• Zero required extras. The tools list on every adapter is always populated — plain dicts in the target framework's field names. The native framework objects (langchain_tools, crewai_tools) are None unless the optional dependency is installed, so importing mcpgram never forces you to install LangChain or CrewAI you're not using.
• Same error shape as the JS SDK. A tool that ran but failed comes back as a normal ExecuteResult(status="error", ...), not an exception — only transport/auth failures raise PlatformApiError.