FauxPilot is an open-source, self-hosted alternative to the GitHub Copilot backend, serving code completions from models running on your own GPU. It has more than 14,600 stars on GitHub and was most recently updated in 2024.
Instead of sending your code to a cloud service, FauxPilot runs Salesforce’s CodeGen models inside NVIDIA’s Triton Inference Server with the FasterTransformer backend, packaged as a Docker Compose setup. The server is designed to stand in for the Copilot endpoint, so editor clients get AI completions while every line of code stays on your machine. Model sizes range from 350M to 16B parameters, and larger models can be split across multiple GPUs if you have the VRAM.
The setup script walks through the heavy lifting: it downloads the CodeGen checkpoint you choose, converts it into the FasterTransformer format Triton serves, and launches the server and client pieces with Docker Compose. Because completion requests are served by a model you control, FauxPilot doubles as a testbed for studying self-hosted code generation โ you can swap model sizes to trade quality against latency and VRAM, and inspect exactly what the model suggests. It began as a community experiment by Brendan Dolan-Gavitt and remains the reference project for anyone asking whether Copilot-style assistance can run entirely offline.
FauxPilot is free and open source under the MIT license. The trade-off is hardware: you need an NVIDIA GPU with enough VRAM for the model size you choose.
Privacy-conscious developers, air-gapped or offline teams, and tinkerers who want Copilot-style completion without a subscription or cloud dependency.