v1.16.0 · self-hosted · MIT licensed

Run your own inference fabric without the usual ops overhead.

Ghostlink turns your LAN into a shared inference layer with discovery, placement, auth, and observability built in — all from a single binary and a simple launcher.

Get started View on GitHub
CPU • GPU • NPU Rust 2021 Windows • Linux • macOS
Ghostlink Studio live cluster
Models Loaded locally
3 active
Workers Networked nodes
7 online
Inference path Balanced across CPU and accelerator tiers
auto-placement enabled
Editor Explain / Fix / Refactor from the workspace
repo-aware context included
Discovery Zero-config Broadcast + mDNS find peers automatically.
Placement Hardware-aware Use each node’s real capability instead of guesses.
Control Authenticated Token auth, rate limiting, and optional PQC-hybrid transport.

What makes Ghostlink feel different

It is built for local-first distributed AI, not a cloud-only serving stack.

Cross-machine inference
Requests can span multiple machines so one model can be larger than any single box.
In-GUI workflows
Explain, fix, refactor, and review code changes directly from the embedded editor.
Local-first design
There is no vendor lock-in dependency on a hosted control plane or cloud runtime.

Launch it in three steps

Clone, build, and start the full local stack with one launcher.

git clone https://github.com/rwilliamspbg-ops/Ghostlink.git
cd Ghostlink
cargo build --release -p ghost-link
./launch.sh

See the experience in motion

The walkthrough below shows the local workflow from first launch to chatting with a loaded model.

Ghostlink Studio walkthrough
The full walkthrough is also available in the repository for deeper setup guidance.

Docs and references

Jump into the architecture, quickstart, benchmarks, and deployment guides.

Architecture
Request flow and cluster layout
Quickstart
Guided setup walkthrough
API Reference
Request and response examples
Benchmarks
Real measured runs
Comparison
vs. vLLM, Ollama, and more
Security
Auth, TLS, and trust boundaries