e4676b5049
The controller itself, as a pure thing with no I/O, so the decisions can be tested away from a connection. It is driven by the receiver's report rather than by this end's view of the path, because this end's view was measured being wrong when it mattered most: 182 ms round trip, zero QUIC packet loss, and a client watching arrivals fall from 1776 to 239 datagrams a second. Some of that is structural -- `send_datagram` evicts the oldest queued datagrams and returns `Ok`, so a sender overrunning a path is told nothing at all. Round-trip time and congestion window remain as a fallback for when the far end has gone quiet, and nothing else. Backing off anchors on measured goodput rather than on a fraction of what was being asked for, which is what makes it quick: at 10 Mbps into a 3 Mbps path the next target is under what actually arrived, so one step does what twenty multiplicative decreases would. The test that matters reproduces the reported failure -- 10 Mbps offered, no frame completing at all -- and requires it corrected within five seconds. Belief and actuation are separate, and that is not tidiness. Folding them together stalls the climb: the step up is a fraction of the ceiling, so past half the ceiling every step falls under the deadband, and a target that only moved when it actuated could never accumulate past that point. The tests caught it. Belief moves every second; saying so is rationed. The first decision is always stated even when nothing changed, because the encoder started at whatever its environment gave it and this end cannot know that matches. Standing down is explicit in three cases: a person set the bitrate by hand, the encoder is under constant quality and has no bitrate to decide, or nothing has been heard and nothing can be seen -- and that last one decays rather than holds, since holding a high target on no evidence is precisely how the original failure sustained itself. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Run your games on a GPU you don't own — or one you do. Nestri puts an interactive workload in a hardware-accelerated virtual machine and streams it to you over QUIC, at a latency that lets you play rather than watch.
Note
This repository is mid-rewrite, and the documentation is behind the code. The guest-side components arrived recently and their docs are thin. Nothing here is stable yet: expect directories to move and interfaces to change. Proper documentation is on the way — issues and questions are welcome in the meantime, and are genuinely useful for deciding what to write first.
Try it now — nesdoctor
One thing here is finished and runs on its own machine, today:
# Linux and macOS
curl -fsSL https://doctor.nestri.io/install.sh | sh
# Windows
powershell -c "irm https://doctor.nestri.io/install.ps1 | iex"
It tells you whether your machine could host games for other people, and
measures the number that actually decides whether streaming a game feels
right — not your download speed, but how much latency your connection adds
when it is busy. A 500 Mbps uplink that queues for 300 ms under load cannot
carry a game; a 25 Mbps one with fq_codel can. Almost nobody has seen their
own figure.
upstream 35 Mbps
latency, idle floor 56 ms
latency, loaded 185 ms
added under load +129 ms grade F
presentation path x11 · bspwm
eDP-1 1920x1200 @ 60 Hz, 8-bit
Vulkan decode h264, h265
It also reads your display out of its EDID — resolution, refresh, colour depth, HDR transfer functions, BT.2020, chroma — and what your hardware can decode. Those decide what is worth sending over the wire, and we would otherwise be guessing from one panel in one room.
It does not stream a game. It is the piece that has to exist before
anything else can, and most machines will come back CLIENT — which is a real
answer, not a failure.
Downloads one binary, verifies its checksum, runs it, deletes it. Installs
nothing, needs no administrator rights, touches no system directory. Nothing is
uploaded: it prints a link, lists exactly what the link contains, and opens it
only if you press Enter. The scripts those URLs serve are
apps/nesdoctor/install/ in this repository, so you
can read them before you run them.
Source and the full story: apps/nesdoctor.
What is here
Two halves that meet over the network and share very little else, plus one thing that runs on your own machine.
The control plane — TypeScript
apps/api |
The public REST API. Identity, teams, machines, games, pairing. |
apps/auth |
A self-hosted OpenAuth issuer — Steam and SSH-key login. |
packages/core |
The domain: every table, every operation, no HTTP. |
packages/auth |
Shared auth types and subjects. |
Postgres for state. Both run on Cloudflare Workers today and as ordinary
containers wherever you like — one handler each, no infrastructure-as-code, and
a Dockerfile in each app. See docs/deploy.md and
docs/dns.md.
The guest — Rust, inside the box
These run inside a virtual machine, beside the game. None of them talk to the control plane.
apps/nescope |
A headless Wayland compositor for one fullscreen client. A lighter answer to the same problem gamescope solves. |
apps/nescapture |
A Vulkan implicit layer. It captures frames from inside the workload's own process and encodes them on the GPU that drew them — no copy out to the CPU and back. |
apps/neswire |
Audio capture and transport. |
apps/neshub |
One connection out of the box. Muxes video, audio, cursor and input into a single QUIC stream to the client. |
crates/nesprotocol |
The wire types they all share, so no two ends can drift apart silently. |
On your own machine — Rust
apps/nesdoctor |
Whether a machine can host a box, and what its connection and display can really do. The first executable form of our host requirements — until it existed, a host was qualified by a human reading a table. Four dependencies; everything that could be done with the standard library is. |
The hypervisor the guest components run under is nesbox,
a separate repository: a micro-VM with a real GPU in it, using virtio-gpu native
context rather than passthrough, so one card can host several boxes at once.
Why a virtual machine
A container shares the host kernel, which makes strong isolation hard and a GPU harder. A micro-VM boots in about as long, isolates properly, and — with native context — gets close to bare-metal graphics. That choice is what makes "many sandboxes, one GPU" possible instead of one tenant per card.
Getting started
bun install
cp .env.example .env # compose reads every credential from here
docker compose up postgres # the database
bun run db:migrate # schema
bun dev # control plane, local Cloudflare runtime
docker compose up --build # or: the whole control plane as containers
cargo build --workspace # guest components
cargo test --workspace
The guest components expect a Linux host with a Wayland-capable GPU stack, and are not much use on their own yet — they are pieces of a box, and the thing that assembles a box is not open yet.
nesdoctor is the exception and needs none of that:
cargo run --release -p nesdoctor
Status
Working: nesdoctor — released, and the only part a stranger can operate
today. The API, auth, the domain model, and the guest components listed above.
Not here yet: the box lifecycle, storage, the edge, and the client. Some of that will open as it is written; some is deliberately closed. What decides which is whether it handles your data — that half is open on principle — or decides our capacity, which is the part we sell.
Contributing
Early, and the ground moves. The two most useful things you can do right now
cost a minute each: run nesdoctor and send the result, because we have
almost no idea what the machines on the other end of this look like; and tell
us where the documentation failed you. Conventional commits; explain why in
the body.