pxpipe

Why it's ranked here

The unusual part is the accounting, not the trick. Savings are measured per request against a token-count probe of the original body, both sides priced under the same observed cache state, and rows where imaging lost money are kept rather than floored. The docs state plainly that the method is lossy, that exact twelve character hex strings were read 13 of 15 times on one model and 0 of 15 on another, and that a verbatim-risk guard is not built yet. Mechanism plus published failure modes is rare. Treat the headline percentage as workload dependent, not a rate you inherit.

What's good

The security posture is written down rather than implied. The server binds loopback by default, and dashboard routes stay loopback-only even when an operator widens the bind address. The hosted worker refuses to serve at all when a key override is configured without a shared secret, and compares that secret by digest instead of by string. Inbound bodies are capped at 16 MiB before allocation. Capturing failing request bodies is off unless explicitly enabled. Config writes go to a temporary file and then rename, at owner-only permissions. One runtime dependency, and the publish step is gated on typecheck plus tests.

Tradeoffs

It is lossy by construction and the misses are quiet. A vision model has no per-glyph confidence, so a bad read arrives as a confident wrong answer rather than an error. Byte-exact values have to stay in text; the bounded factsheet preserves up to 96 recognized precision-critical tokens, not every identifier. The gain swings with the workload: dense code and tool output win, sparse prose loses money, and a profitability gate is the only thing standing between you and the bad case. Two model families are enabled by default. Everything else, including several strong readers, is opt-in.

How to use it well

For anyone running a long-lived coding agent against a metered API, where the same system prompt and tool docs are resent every turn. It earns a slot when sessions run long, context is dense, and the bill is real. Send byte-exact work, hashes, identifiers, secrets, to a passthrough model or keep it in the recent text tail. There is an offline mode that renders files or a diff to pages for image-upload clients, with no proxy involved. It does not compress model output, does not add caching of its own, and does nothing for answer quality.

Observed

License

MIT

Language

TypeScript, ESM only

Install surface

npm package exposing one binary, runnable through npx without installing

Interfaces

Local HTTP proxy, loopback dashboard, offline export and stats subcommands, and typed library entry points

Runtimes

Node http server and Cloudflare Workers share the same core proxy

Runtime dependencies

One, a tokenizer library; everything else is a dev dependency

Default binding

Loopback address and port 47821; dashboard routes stay loopback-only even if the bind is widened

Request ceiling

16 MiB inbound body limit enforced before allocation, operator-overridable

Test surface

Vitest suite plus typecheck and build gated in the prepublish step

Security documentation

docs/SECURITY_MODEL.md gives assets, trust boundaries, a threat and control table, and out-of-scope risks