Agent frameworks

Any MCP client can start frankensurf-mcp (Add to your agent). For frameworks that take Python tools, frankensurf.agent_tools gives you two ready-made ones:

• frankensurf_read: a page’s text (or the article, with main), its title, a quality grade, which tool got it, and a trace_id to pass back as try_harder_than when the page wasn’t right. With a JSON Schema, the fields filled from the page’s own data (Structured fields).
• frankensurf_search: checked results (url, title, snippet, published), with recency, site and vertical.

Policy settings pass straight through, for example langchain_tools(allow_paid_fallbacks=True, max_cost_usd=0.02). Text is capped at 20,000 characters (max_text).

LangChain

pip install "frankensurf[langchain]"

from frankensurf.agent_tools import langchain_tools
from langgraph.prebuilt import create_react_agent

agent = create_react_agent(model, tools=langchain_tools())

Each tool works sync (invoke) and async (ainvoke).

OpenAI Agents SDK

pip install "frankensurf[openai-agents]"

from agents import Agent, Runner
from frankensurf.agent_tools import openai_agents_tools

agent = Agent(name="researcher", instructions="Use the web tools to answer.", tools=openai_agents_tools())
result = await Runner.run(agent, "What changed in the latest ferry timetable?")

The SDK’s strict tool schemas can’t describe a free-form object, so here the schema is passed as JSON text: extract_schema_json.

CrewAI

pip install "frankensurf[crewai]"

from crewai import Agent
from frankensurf.agent_tools import crewai_tools

researcher = Agent(role="Researcher", goal="Find and read sources", backstory="...", tools=crewai_tools())

CrewAI supports Python up to 3.13.

Claude Code plugin

Install the package (pip install "frankensurf[mcp]", then frankensurf install), then:

/plugin marketplace add yail259/frankensurf
/plugin install frankensurf@frankensurf

The plugin starts the MCP server and adds the skill that teaches Claude when to read, search, try harder, or hand a wall to you.