RogerAI

             ·   ˙   ·
           ˙ ((  •  )) ˙                   ┌──────────────┐
              \(   )/                      │ ((•)) ON AIR │
               │ R │         ▟█▙           └──────────────┘
   ┌───────────┴───┴──────────█───────────────────────────┐
   │  ◉ ch 01 · on band gpt-oss-20b · 42 tok/s   ▂ ▄ ▆ █  │
   ╰┬────────────────────────────────────────────────────┬╯
    ▔                                                    ▔

        █▀▀█ █▀▀█ █▀▀▀ █▀▀▀ █▀▀█     █▀▀█ ▀█▀  ▄
        █▀▀▄ █  █ █ ▀█ █▀▀  █▀▀▄  ▄  █▀▀█  █   █
        ▀  ▀ ▀▀▀▀ ▀▀▀▀ ▀▀▀▀ ▀  ▀     ▀  ▀ ▀▀▀  ▀
             borrow a GPU, pay by the token · roger that.

A two-way radio for GPUs. RogerAI is a marketplace for crowd-sourced, self-hosted LLMs: people run open models on their own GPUs and go "on air"; you tune in and pay per token. Every token carries a model-lineage record - a receipt signed by the provider and counter-signed by the broker, tracing the response back to the model that produced it. Owners monetize idle hardware; users get cheap, diverse access.

curl -fsSL https://rogerai.fm/install.sh | sh

Or via Homebrew (macOS + Linux). Homebrew 6+ makes you trust a third-party tap once, so it's a single line:

brew trust rogerai-fyi/homebrew-tap && brew install rogerai-fyi/homebrew-tap/roger

Then just run roger for the interactive radio (browse stations, tune in, test, copy the endpoint; rogerai is the legacy alias).

Use it

roger                         # interactive TUI: browse → connect → test → copy endpoint
roger search                  # list models (cheapest now first; shows tok/s, ◆ confidential, FREE)
roger use <model>             # open a local OpenAI-compatible endpoint that relays via the broker
roger voices                  # browse voices on air (TTS/STT stations, sample previews)
roger say --voice <v> "hi"    # text-to-speech through a voice station, played locally
roger balance                 # wallet credits
roger topup 10                # buy credits (Stripe)
roger remote                  # your private BASE STATION: continue an agent session from anywhere

Inside the agent ([0] AGENT), /remote-control puts the live session on your private Base Station so you can pick it up from another terminal (roger remote attach <code>), the web console, or the app - private to your account, one-time link codes for a phone. Tools keep running on the host and still ask before anything mutating.

roger use (or the TUI's "tune in") exposes http://127.0.0.1:4141/v1 with an API key - point any OpenAI-compatible tool at it.

Share your GPU (become a provider)

The node dials OUT and long-polls the broker for jobs - no inbound ports, no tunnel dependency. Behind any NAT, one command:

roger share                                    # auto-detects your local model, starts earning
# options: --price-in/--price-out, --free-window 03:00-03:30, --schedule '<time-of-use JSON>',
#          --confidential (TEE-attested), --upstream <your OpenAI endpoint>

Bring your preferred model host. Compatibility is determined by the served API, not the trainer, optimizer, quantizer, model format, or vendor. RogerAI auto-detects common local hosts, including Ollama, llama.cpp, LM Studio, Unsloth Studio, vLLM, Osaurus, Jan, and LiteLLM. LocalAI, TGI, SGLang, KoboldCpp, text-generation-webui, and any other host implementing the supported OpenAI-compatible API work through --upstream (and --upstream-key when needed). An Unsloth-trained or Unsloth-quantized model also works when served by any compatible host.

Named integrations make discovery easier; they are not a vendor allowlist. See hosting compatibility and current gaps for the exact upstream contract. On the other side, roger use gives clients an OpenAI-compatible endpoint too.

How it works

rogerai ──► broker (broker.rogerai.fm) ──► your node ──► your local model
 discover     registry · wallet · relay      dials out      (Ollama/llama.cpp/LM Studio/vLLM/Osaurus/…)
 use/topup    match · meter · co-sign         serve+sign

The broker is the only public component and is content-blind (it stores token counts and signed receipts, never prompts). It's an OpenAI-compatible relay - see the served spec at /openapi.yaml.

• Per-token pricing with a 24h price-lock, free and time-of-use windows.
• Voice bands - TTS/STT stations alongside chat: share a Kokoro-style speech server with roger share, browse with roger voices, attributed on air as @station/slug.
• Lineage receipts - hash-chained, dual-signed (internal/protocol).
• Privacy - identity pseudonymized to providers; a confidential (TEE) tier for sensitive work.
• Routing constraints - price, measured throughput (tok/s), confidential-only.

Run it locally

make build
make demo     # broker + a node + a request, end to end (needs a local OpenAI endpoint)
go test ./...

Docs

• BROKER-SPEC.md - the open broker spec (anyone can self-host/federate)
• cmd/rogerai-broker/openapi.yaml - OpenAPI 3.1 (also served at /openapi.yaml)
• VERIFICATION.md · PRIVACY.md · DEPLOY.md · STRIPE.md

License

Two licenses, side by side (LICENSING.md has the exact file list):

• The node-agent protocol and the usage-receipt SDK are Apache-2.0. Anyone can implement a compatible node, or verify what they were charged, without asking. An interoperability surface only counts as open if it is actually open.
• Everything else is PolyForm Perimeter 1.0.0 - source-available, free for any non-competing use. You can self-host your own broker for your own community; you can't run it as a competing commercial marketplace service. See the LICENSE for the exact grant.