FauxPilot
Self-hosted GitHub Copilot clone — run Salesforce CodeGen models on your own GPUs behind the Copilot API.
About
FauxPilot is an open-source (MIT) attempt to build a locally hosted GitHub Copilot. It runs Salesforce CodeGen models inside NVIDIA's Triton Inference Server with a FasterTransformer backend, and mimics the Copilot API surface so standard Copilot clients in VS Code or Neovim can connect to it. The pitch is simple: Copilot-like completions on your own infrastructure, with zero code leaving your network — attractive for privacy-sensitive teams and tinkerers.
The honest picture: it is a 2022-era project with the last meaningful commits in early 2024, and the underlying CodeGen models are far behind modern coding models. Setup is genuinely heavyweight — Docker, nvidia-docker, and an NVIDIA GPU with serious VRAM — and completion quality is a fraction of today's hosted assistants. Consider it a reference architecture for self-hosted inference more than a daily driver.
What it can do
- Copilot-API-compatible completion server
- Salesforce CodeGen models via Triton Inference Server
- Works with standard Copilot VS Code/Neovim clients
- Multi-GPU model splitting
- No code leaves your infrastructure
Setup
Clone the repo, run the setup script to pick and download a model (needs an NVIDIA GPU, Docker, nvidia-docker), then point a Copilot-compatible client at your local server.
Prerequisites
Docker with nvidia-docker, an NVIDIA GPU with enough VRAM, and comfort with model conversion tooling.
Use cases
Works with
Trust score — 6/10
Scored 6/10 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 57; Trending score 58/100 from researcher assessment.
★ 14,733 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.