Surfil
// compare

Compared on architecture, not adjectives.

There are three real alternatives to Surfil: build a proxy, stitch point tools, or route through a cloud gateway. Each is legitimate - here's the honest capability map.

surfil · bench - acme/api
Quality on your repo
Bar chart: Opus 96, Sonnet 89, GPT 81
Opus - 42% fewer fixes / taskrecommended
The alternatives

What you'd actually be choosing between

Steelmanned on purpose - a comparison table only means something if the other columns are described fairly.

DIY proxy

A mitmproxy or custom gateway you build and maintain. Total control, but the proxy sees every prompt, you own its security surface, and it becomes an unfunded internal product.

Point tools

One cost tracker, one secret scanner, one memory plugin - per agent. Each is fine alone; together they're N configs, N failure modes, and no shared source of truth.

Cloud gateway

Route agent traffic through a vendor's cloud. Easy to adopt - but the gateway reads everything, cost scales with traffic, and leaving means unpicking your workflow.

The matrix

Capability by capability

Every row is an architectural fact - checkable from how each approach is built, no benchmarks required.

CapabilitySurfilDIY proxyPoint toolsCloud gateway
Source never leaves device✓ By construction✗ Proxies see prompts~ Varies by tool✗ Cloud gateway sees all
Signed, offline-verifiable output✓ Ed25519, no account~ Vendor attestation only
One interception layer✓ One-Layer rule✗ You maintain it✗ One per tool
Cost flat per user✓ Device does the work~ Your infra cost✗ Scales with traffic
Portable memory (export)✓ .surfilpkg~✗ Lock-in
Honest, measured cost✓ measured + signedn/a~ Marketing rates~ Marketing rates
Full support
~ Partial / varies
Not supported
The other direction

When Surfil is the wrong choice

A comparison page that never concedes anything is an ad. Three cases where an alternative genuinely wins.

You need traffic inspected in a central cloud

Some org policies require a cloud chokepoint that reads content. Zero-trace is the opposite architecture - a gateway fits that mandate; we don't.

You want one metric, once

If all you need is a spend counter for a single agent, a point tool is less to install. Surfil earns its place when cost, safety and memory stack up.

You enjoy owning the proxy

Some platform teams want interception as their own product. Genuinely fine - our One-Layer rule just means Surfil won't sit behind it.

Why no vendor names

The missing column you might expect

No competitor names, no “X% better than BrandY” - here's the reasoning, not just the rule.

Named-competitor tables need fixed numbers to look decisive, and fixed numbers about someone else's moving product are exactly the unmeasured claims our honest-claims policy forbids. Architecture classes don't change week to week; vendor feature lists do. So we compare against the class - and let your own measured receipt settle the rest.

We stand behind no savings percentage at all. Every cost number is measured on your traffic before it is shown, and signed after.

Settle it with a receipt, not a table

One Core install. One week of your real traffic beats every row above - and uninstalling restores everything byte-for-byte if we lose.