@piedpiperrh/host

Host LLM inference for the Pied Piper network from your terminal and get paid in credits for every request you serve. No browser tab required — inference runs through your local Ollama.

Prerequisites

1. Ollama installed and running (ollama serve happens automatically on most installs).
2. At least one supported model pulled:

   ollama pull llama3.2:3b     # small, popular, good starter
   ollama pull qwen2.5:7b      # if you have ~6 GB of VRAM

3. A Pied Piper account and an API key (dashboard → Settings → API keys).

Run

npx @piedpiperrh/host --key lg_pk_your_key_here

or install globally:

npm install -g @piedpiperrh/host
piedpiper-host --key lg_pk_your_key_here

The CLI detects which of your installed Ollama models exist in the Pied Piper catalog, registers them, and starts serving jobs. Leave it running; stop with Ctrl-C (it deregisters cleanly).

Options

| Flag | Meaning | | --- | --- | | --key, -k | API key (lg_pk_…); or set PIEDPIPER_API_KEY | | --price <n> | Credits you charge per request (default 1) | | --title <text> | Marketplace listing title; also makes your host rentable/pinnable | | --vram <mb> | Advertised VRAM cap (default: sized from your largest served model) | | --ollama <url> | Ollama endpoint (default http://127.0.0.1:11434) | | --url <url> | Platform URL, for previews/self-hosted deployments |

How you get paid

Each completed request transfers credits (your --price, ×2–10 for priority requests) from the renter's balance to yours. Check earnings in the dashboard; credits can be withdrawn on-chain from the Credits page.

Supported models

The network catalogs models by web-llm id; this CLI serves the equivalent Ollama weights. Currently mapped: Llama 3.1/3.2, Qwen 2.5 (+Coder), Gemma 2, Phi 3.5, Mistral 7B, SmolLM2, DeepSeek-R1 distills, Hermes 3, TinyLlama, plus the vision models llama3.2-vision, llava, qwen2.5vl, and moondream. Anything else installed in Ollama is ignored (the catalog and the CLI's model map have to agree on identity — PRs welcome).

Tool calling & vision

Console hosts are the network's capability tier:

• If any of your served models supports function calling (Llama 3.x, Qwen 2.5, Mistral, Hermes 3), your host advertises tools and receives OpenAI-style tool-calling jobs.
• If you serve a vision model (ollama pull llava or llama3.2-vision), your host advertises vision and receives image requests — images arrive inline in the job payload, nothing is fetched from the internet.

Browser hosts can't do either, so these requests route exclusively to console hosts — usually a less crowded, better-paying lane.

Notes

• Your machine only receives job payloads (messages + sampling params) and streams tokens back; prompts are scrubbed from the platform DB after each job settles.
• Keep the process alive to stay in the dispatch pool; a missed heartbeat for ~5 minutes ends your session server-side.