How it fits together

Your agent makes one call. Frankensurf checks your settings, works out who the request should run as, picks the cheapest way to fetch the page, checks the result and hands it back with a receipt.

  Your agent (any model)  ── Python · command line · MCP ──┐
                                                            ▼
 ┌────────────────────────────── Frankensurf ─────────────────────────────┐
 │  your settings ─▶ who to run as ─▶ router ◀── memory of what worked    │
 │                                       │                                │
 │        add-ons can step in before a request, after it, or on failure   │
 │                                       ▼                                │
 │  plugins: browsers · scrapers · search · site parsers · login storage  │
 │                                       ▼                                │
 │  saved pages · cache · receipts · traces · benchmark                   │
 └────────────────────────────────────────────────────────────────────────┘
                  │
                  ▼
        the web, through plain requests, browsers, stealth browsers,
        unblocking services and your own logged-in Chrome

The rules it follows

Get the job done. Getting the right page comes first. Speed and cost come second.
No hidden limits. Every limit is a setting with a default. You can change any of them.
Everything plugs in. Browsers, scrapers, search engines and site parsers are all plugins.
Your secrets stay yours. Passwords and cookies never reach the agent or the AI model.
Say what you know. Every result has a receipt. If something is unknown, it says so.
Measure, don’t guess. The benchmark decides which tools become defaults.

The parts

Part	What it does	Today
Settings	Every limit, in one place	partly some limits are still fixed in code
Router	Picks how to fetch each page	fixed order
Memory	Remembers what worked on each site	planned
Measurements	Success, speed and cost per tool	recorded
Plugins	Lets you add any tool	planned tools are built in for now
Logins	Fetch pages as you	working on your own machine
Receipts	Show how each result was fetched	working

Local first

Everything runs on your machine. You don’t need an account or any hosted service. If you add a paid service yourself, like an unblocking API, it’s just another plugin.

A hosted Frankensurf service may come later for extra scale and team features. It will never control your local setup. It can suggest things, and your machine decides.

Before 1.0

Frankensurf hits 1.0 when it passes all six of these checks on the benchmark.

Check	Passes when
Better than any single tool	It succeeds at least as often as the best single tool on every kind of site, and fails at most half as often overall.
Cheaper	It sends at most half the tokens of a plain browser snapshot, and costs no more per success than the best single tool.
Learns	A second run is at least 30% faster, with almost no wasted attempts.
No silent mistakes	Fewer than 1 in 200 “successes” contain the wrong content.
Works with any model	Three or more AI apps or models get the same results.
Quick to start	A new machine gets its first result within five minutes of opening the README.

The benchmark

The benchmark (FrankenBench) runs the same list of pages through each tool on its own, through a plain browser, and through Frankensurf. Every page has checks for the content it should contain, so “success” means the right content, not just a page that loaded. It records success, mistakes, tokens, cost, speed and attempts, and its results feed the website.

What’s done and what’s next

	Stage	Status
A	Fetch pages locally, with receipts	done
B	Logins and safety rules	partly
H	The benchmark	next
P	Plugins, and turning fixed limits into settings	next
M	Remembering what works	planned
C	Cheaper ways to fetch: markdown pages, signed requests, shorter outputs	planned
D	Running on your machine from elsewhere, sharing logins safely	planned
E	Auto-written site parsers, shared learning, hosted service	planned