protogrid (Python)

Client for the protogrid registry: find MCP servers by intent, get a machine-readable connection block, and connect with no human in the loop wherever the server allows it.

import os
from protogrid import ProtogridClient, find_connectable, open_session

registry = ProtogridClient()  # public registry; ProtogridClient("http://localhost:8080") for a local stack
found = find_connectable(registry, "send an email", secrets=os.environ)
async with open_session(found.connection, os.environ) as session:   # mcp.ClientSession, initialized
    tools = await session.list_tools()

• ProtogridClient / AsyncProtogridClient: search, get_server, list_tools, list_all_tools, get_connection, and check / get_check for an on-demand check of any remote MCP server URL (quality and Claude and OpenAI directory readiness).
• No key needed. A free key from https://protogrid.dev/account raises the limits (60 requests per minute, 5,000 per day); the client reads PROTOGRID_API_KEY, or pass ProtogridClient(api_key=...).
• Connection blocks carry ${NAME} placeholders; substitute_secrets fills them from your own store. The registry never sees secret values.
• connection_class: R0 remote, no auth · R1 remote, static secret you hold · R2 remote OAuth (one consent) · L0 local package (allow_local=True) · unknown.
• R2: open_session(conn, oauth=OAuthOptions(store, consent)) runs the one-time consent (loopback_consent or manual_consent) through the official SDK's OAuth provider and keeps tokens in your TokenStore; later runs need no human.
• PydanticAI: to_pydantic_ai(conn, secrets) returns an MCPToolset. to_fastmcp_transport and to_mcp_servers are pure formatters.
• Install: pip install protogrid-sdk (imported as protogrid). Extras: protogrid-sdk[mcp] for open_session, protogrid-sdk[pydantic-ai] for the toolset.

Examples: examples/find_and_call.py (search → connect → call a tool) and examples/pydantic_ai_agent.py (agent with a registry-found toolset, no LLM key needed).