Tendril turns one multi-GPU machine into several plug-and-play gaming stations — each a Windows or SteamOS VM with a real GPU passed through. It's a self-healing Fedora bootc appliance you drive from a web UI or an on-screen console; it handles IOMMU, VFIO, driver binding, and VM setup so you don't have to.
🎮 Live demo: click around the real web UI (read-only — actions are disabled) at demo.onetick.ninja. No install needed.
Development happens on a self-hosted Gitea instance; github.com/sudolulo/tendril is the
public mirror, and issues and discussion are welcome here. That's also why the bootc commands
below pull from git.onetick.ninja — that's the project's container registry, and it is
anonymously pullable, so the commands work as written with no account.
One powerful box with several GPUs → several independent gaming setups at once. Two people gaming on one tower, a handful of Steam Machines driven from a closet server, a Windows VM for the games that need it next to a SteamOS VM for everything else — up to a room of machines imaging themselves over the network.
- Passthrough-first: N GPUs → N independent stations (the reliable path on consumer hardware).
- Self-healing host: atomic bootc images with greenboot auto-rollback — a bad update can't brick the box.
- Own libvirt orchestrator: full control of passthrough, CPU pinning, and Secure Boot + TPM (for Windows 11).
- One edition, AGPL: every feature is open source on every channel.
Stations. Create a station from the browser — pick the OS, GPU, and account — and Tendril builds the disk, the answer-file/kickstart seed, and the VM, then installs the guest completely unattended (Windows 11 past the virtio and Microsoft-account walls, or a SteamOS-style Bazzite image) and boots from disk. Install media is auto-fetched and checksum-verified; a live in-browser noVNC console and install progress come free. Low-latency mode pins vCPUs to dedicated host cores and uses hugepages so frame times don't jitter.
Day-2. Snapshots (restore points before risky updates), golden images (capture an installed station, clone it instantly copy-on-write, push it to stations fleet-wide), optional persistent data volumes that survive OS reinstalls and reimages (with backup/restore to the shared store), USB hot-plug and seats (named USB device groups), and remote play — stations run Sunshine, any device runs Moonlight. Stations stay editable after creation (vCPUs, memory, GPU), show their real disk usage with one-click compact, run on daily schedules, and a kiosk mode resets a station to its golden image on every start — walk-up-and-play for venues.
The host. HTTPS by default with cert management in the UI, admin + read-only viewer logins plus optional named users and API tokens — all feeding an audit log that names who did what — live logs, webhook/ntfy notifications for the events that matter, Prometheus metrics, one-click settings backup/restore, configurable networking with a 60-second test-and-revert safety net, NFS/SMB shared storage, and one-click OS updates (staged by bootc, applied on reboot, auto-rolled-back by greenboot if the new image doesn't boot healthy).
vGPU (🧪 needs real-hardware validation). Split one GPU across several stations (mdev or
SR-IOV — see the supported-card list). Stage NVIDIA's licensed host .run in the
UI and Tendril builds the driver image variant on the box; the matching guest driver and
licensing are then fully automatic — new vGPU stations come up licensed with zero manual steps,
and a data-preserving re-split changes a station's slice without touching its disk.
Fleet (🧪 needs real-hardware validation). Manage several boxes from any one of them: join codes (paste one string — trust, mTLS, and mutual membership included), mDNS discovery of nearby nodes, GPU-aware placement, control of any node's stations from any UI including the cross-node console, golden-image distribute, human-confirmed cold re-home of an image-backed station off a dead node, staggered fleet OS updates (one node at a time, waiting for each to come back healthy), and PXE room provisioning — netboot a rack of bare-metal PCs straight into the unattended installer. (docs/FEDERATION.md)
Games at scale. Bake games into a golden image, or share one Steam library over virtio-fs — install once, play from many (docs/STEAM-GAMES.md).
Drive it all from the web UI, the console on the attached display, or the CLIs (docs/CLI.md).
Easiest: download the installer ISO from https://dl.onetick.ninja/, verify it against
SHA256SUMS, flash it to a USB stick, boot the target, and install. Then open https://<host-ip>
(self-signed cert — the UI can install a real one), set an admin password, and create your first
station.
Or deploy the published image with bootc:
sudo bootc switch git.onetick.ninja/flan/tendril:latest && sudo rebootThree channels are published — rolling :dev, per-release :latest, and validated :stable —
plus matching installer ISOs; releases are cosign-signed. See docs/CHANNELS.md.
Prerequisite: enable VT-d (Intel) or AMD-Vi / IOMMU (AMD) in your motherboard's firmware — no software can turn this on for you. Building the image yourself and the full walkthrough: docs/INSTALL.md. Still pre-1.0; expect rough edges.
| Doc | What's in it |
|---|---|
| docs/INSTALL.md | Install from ISO, build the image yourself, first station (web + CLI) |
| docs/CHANNELS.md | dev / latest / stable channels, what "stable" promises, signing |
| docs/FEDERATION.md | Fleets: join codes, mTLS, placement, re-home, PXE — and the design rationale |
| docs/VGPU.md | GPU splitting: supported cards, driver staging, automatic guest driver + licensing |
| docs/STEAM-GAMES.md | Getting large game libraries onto many stations |
| docs/CLI.md | The command-line tools behind the UI |
| docs/HARDWARE-TESTING.md | The real-hardware validation checklist (testers start here) |
| docs/VERSIONING.md | SemVer policy and how versions are pinned |
| docs/PLAN.md | The original design document (kept for the record, with supersessions noted) |
Everything in "What works today" is implemented and continuously reviewed; the 🧪 items (vGPU, fleet/PXE, remote play) are code-complete but awaiting validation on real multi-GPU hardware — docs/HARDWARE-TESTING.md is the checklist if you can help. Multi-machine is deliberately federation, not clustering: independent self-managing nodes, no consensus, no live migration (a station is its GPU) — the reasoning is in docs/FEDERATION.md.
Trunk-based on dev, Conventional Commits, changelog per change, and a one-time
CLA with your first PR — see CONTRIBUTING.md. Security issues go to
SECURITY.md, not the issue tracker.
Portions of this project — including design documents and code — were produced with the assistance of AI tools. All output is reviewed by human maintainers before it lands. See NOTICE.
AGPL-3.0-only — one edition, every feature open source. Use it freely, commercially included: running stations you charge for carries no obligations. If you modify Tendril and offer it to others over a network, the AGPL asks you to share those changes; organizations that need different terms can get a commercial license — see LICENSING.md.