Netserve is an experiment in representing supported server behavior as neural-network weights. It is not an LLM answering HTTP requests, and the compiled protocol graphs do not learn online. A compiler turns bounded operations into exact threshold networks that a native Goose evaluator executes.
Try the no-signup playground | Repository The graphs implement route recognition, static response generation and supported handlers. The host owns sockets, TLS, authentication, storage, deadlines and queue limits. Keeping that boundary explicit avoids treating cryptography or authorization as approximate model predictions.
A threshold neuron sums weighted inputs and a bias, then emits a bit according to whether the sum reaches zero. Boolean operations can be represented with these neurons and composed into exact circuits. Netserve supplies input bytes and state bits to a graph; its outputs describe response bytes and next-state bits. The supported language is restricted: this is not an arbitrary JavaScript-to-network compiler.
A counter computes its next eight-bit state and JSON response, wrapping from 255 to zero. An echo graph preserves up to 64 UTF-8 bytes. A codec packs up to 32 characters from a 64-symbol alphabet into six bits per character, plus a length byte. Both encoding and decoding execute as networks. The sample shown in the demo goes from 30 to 24 bytes and restores exactly. Short inputs can grow, unsupported symbols are rejected, and this is not a learned universal compressor.
The browser uses a JavaScript threshold evaluator. The downloadable server uses native Goose. A parity test compares the browser evaluator with Goose across 359 cases, including counter wraparound. Platform CI builds and tests Windows x64, Linux x64, macOS ARM64 and macOS x64. The source includes the tests so the claims can be inspected and rerun.
The local STUN/TURN experiment uses neural request recognition, protocol decisions and encoding, with host-enforced authentication and bounds. Its TURN peer policy permits only other live local relay allocations. A WebRTC test checks that both selected candidates are relay candidates and exchanges a message. This bounded local subset is not a production coturn replacement, a public relay, or an anonymity service.
The project's 90.32% neural coverage is an equally weighted feature inventory, 28 neural units out of 31. It includes the default app routes and supported protocol/demo behaviors; secure host services, browser UI, build tools and historical experiments are excluded. Conventional HTTP framing, TURN attribute/peer parsing and response/inbound-channel assembly are counted as host behavior. This is not an independent certification or a measure of source code, runtime instructions, CPU work or memory. The complete inventory and measurement script are linked on the playground.
The interesting question is where exact compiled networks are useful enough to justify their overhead. Netserve does not establish a performance advantage over ordinary web servers. It provides a working, bounded experiment whose representation, host boundary and behavior can be examined.
Try the browser playground without signing in, or download a tested native package. Node.js 24+ is required for the server; setup instructions and local-security guidance are in BUILDING.md. Downloads | Build instructions | Coverage inventory
Disclosure: This article was drafted by an AI assistant using the project's source and test results. AI tools also assisted development and testing.