Frankensurf

Frankensurf is an open-source Python tool from the GitHub user yail259 that gives AI agents a single call for reading and searching the web. It stitches together dozens of existing web tools, starts with the cheapest, checks that the returned page is real, and escalates only when a site pushes back. The latest release listed on GitHub is v0.51.0, and the README describes the project as an alpha.

What It Is

Frankensurf sits in front of fetchers, browsers, stealth browsers, hosted browser services, paid scraping APIs, and search sources. An agent calls read(url) and Frankensurf tries plain HTTP, Jina Reader, a local browser, and free stealth browsers first, then hosted browsers and paid APIs only if the user allows them. A challenge page, login wall, or search page whose results never loaded counts as a failure rather than a success. Results come back as raw page data (text or markdown, JSON-LD, the page's own JSON, images) plus a receipt showing which tools ran, how long each took, what it cost, and why it stopped.

How the escalation works

Every fetch method is a rung on a ladder, running from content negotiation and light fetch through managed browsers, stealth with signed identity, and unblocker networks to a human handoff for CAPTCHAs and 2FA. If a result is wrong, the agent can retry with a trace ID and Frankensurf skips tools it already tried. Paid tools are opt-in, and a per-read cost cap can be set. Signed-in pages are read through saved profiles, so the model sees profile names rather than cookies or passwords.

Search, extraction and site knowledge

Search uses free sources first, including a local SearXNG engine installed without Docker, and checks each result for topic, page type, and date window. Paid sources such as Exa, Brave, Tavily, and Serper join only when allowed. Callers can request structured fields via a JSON Schema, map and crawl sites, watch feeds and sitemaps for new pages, and save what the agent learns about a site as reusable site modules. Opt-in sharing of which tool worked on which site is also supported.

Where it plugs in

It runs as an MCP server, a Claude Code plugin, a Python library, a CLI, and tool sets for LangChain, the OpenAI Agents SDK, and CrewAI. A Firecrawl-compatible server mode lets tools that accept a Firecrawl URL point at it. Providers and search sources are plugins, so custom tools can be added and must be explicitly trusted by name. The project publishes its own benchmark results comparing it against single tools on 150 sites; these are vendor-run figures.