roger

What RogerAI is

RogerAI describes itself as a two-way radio for GPUs. It is a marketplace for crowd-sourced, self-hosted large language models where people run open models on their own GPUs and go on air, and other users tune in and pay per token. The README frames the exchange as borrow a GPU, pay by the token. Every token carries a model-lineage record, which is a receipt signed by the provider and counter-signed by the broker, tracing a response back to the model that produced it. Owners monetize idle hardware while users get cheap and diverse model access. Installation is offered two ways: a curl piped install script from rogerai.fm, or Homebrew on macOS and Linux, where Homebrew 6 or newer lets a user trust a third-party tap once and then install with a single line. After install, running roger opens an interactive terminal interface to browse stations, tune in, test, and copy an endpoint, with rogerai kept as a legacy alias. The project is written in Go, and the README shows make build, make demo, and go test as the local development commands, where the demo brings up a broker, a node, and a request end to end and needs a local OpenAI endpoint to function.

How a provider shares a GPU

The provider side is designed to work behind any network address translation without inbound ports or a tunnel dependency. The README explains that the node dials out and long-polls the broker for jobs, so a single command starts earning after the local model is auto-detected. Options include price in and price out, a free window such as 03:00 to 03:30, a schedule given as time-of-use JSON, a confidential flag for trusted execution environment attested runs, and an upstream pointing at an OpenAI endpoint. Compatibility is determined by the served API rather than by the trainer, optimizer, quantizer, model format, or vendor, which means a wide range of local hosts work. The README names Ollama, llama.cpp, LM Studio, Unsloth Studio, vLLM, Osaurus, Jan, and LiteLLM as auto-detected, and says LocalAI, TGI, SGLang, KoboldCpp, and text-generation-webui work through the upstream flag. An Unsloth-trained or Unsloth-quantized model also works when served by any compatible host. The broker is the only public component and is described as content-blind, storing token counts and signed receipts while never seeing prompts, which keeps user text out of the relay path. The broker is the only public component and is described as content-blind: it stores token counts and signed receipts but never prompts, and it acts as an OpenAI-compatible relay with a served OpenAPI spec.

Client commands and privacy model

For users, the README lists a set of client commands built around the roger binary. roger search lists models cheapest first and shows tokens per second, a confidential marker, and free availability. roger use opens a local OpenAI-compatible endpoint that relays through the broker, and roger voices browses text-to-speech and speech-to-text stations with sample previews. roger say sends text through a voice station and plays it locally, while roger balance and roger topup manage wallet credits with Stripe. There is also a remote command that puts a live agent session on a private Base Station so it can be picked up from another terminal, the web console, or the app, with one-time link codes for a phone and tools that keep running on the host and still ask before anything mutating. Privacy is handled by pseudonymizing identity to providers and offering a confidential trusted execution environment tier for sensitive work. Routing constraints let a user filter by price, measured throughput in tokens per second, and confidential-only requirements. The pricing model uses per-token billing with a 24 hour price lock plus free and time-of-use windows, which gives both sides predictable economics for short sessions. The confidential flag offers a trusted execution environment tier for sensitive work, while identity is pseudonymized to providers so a node operator does not see who sent a prompt.

Editorial conclusion

RogerAI is written in Go, was last updated on GitHub on 2026-08-24, and carries a dual license where the node-agent protocol and receipt SDK are Apache-2.0 while the rest is PolyForm Perimeter 1.0.0.