EditorAI Platinum — Distributed Inference Network

Free distributed computing system that lets users donate GPU/CPU resources — or any AI provider endpoint they have access to — to run models for EditorAI and other projects.

Thank you VLT.gg for the server making this possible!

📋 For EditorAI Users

1. Change your Ollama URL in EditorAI settings to http://sn-1.vltgg.net:21801 (deprecated — now, simply enable Platinum)
2. Set the Ollama model to any supported model (command listed below)

To list all models, run:

curl http://sn-1.vltgg.net:21801/api/tags

The API is Ollama-compatible and unchanged:

curl http://sn-1.vltgg.net:21801/api/generate \
  -d '{"model":"best","prompt":"hello","stream":true}'

stream: true streams tokens as NDJSON, one chunk per piece of text, ending with {"done": true} — the same shape real Ollama uses, so any Ollama client that accumulates chunks works. Omit stream (or send false) to get a single buffered JSON object instead.

---

👥 For Donors

Install Python from https://python.org, then pip install -r requirements.txt.

A worker can host as many providers as you like. A provider is any URL that speaks one of three API styles:

style	request	typical providers
openai	POST /v1/chat/completions	OpenAI, DeepSeek, llama.cpp, vLLM, LM Studio, OpenRouter, Groq, …
ollama	POST /api/generate	a local or remote Ollama server
anthropic	POST /v1/messages	Claude

For each provider you either map your own tags onto specific upstream models, or let the worker autodetect every model the endpoint reports. Tags are what the network advertises, so anything you map shows up in /api/tags.

Presets

python worker/client.py provider presets

Built-in presets (each still needs your API key when the upstream requires one): ollama, llamacpp, lmstudio, vllm, openai, anthropic, deepseek, openrouter, groq, mistral, together, fireworks, cerebras, xai, gemini, zhipu, perplexity, plus custom for anything else.

Adding providers

The worker is a single self-contained file (worker/client.py). The fastest way to add a bunch of endpoints at once is add — one URL per line, with an optional key, models and headers; the API style and model list are detected automatically:

# interactive — paste a list of endpoints, blank line when done
python worker/client.py add

# ...or from a file
python worker/client.py add --file endpoints.txt

# ...or straight on the command line (URL first, then optional key/models/headers)
python worker/client.py add \
    "https://api.deepseek.com sk-abc" \
    "https://api.openai.com sk-def models=fast=gpt-4o-mini,smart=o3" \
    "https://gw.example.com/v1 sk-xyz header='X-Org: acme' header='X-Env: prod'"

add is idempotent — re-pasting a URL updates it instead of duplicating it. models= is only required for endpoints that can't be autodetected (the tool tells you which ones those are). Fields: a bare token is the key, or use key=, key-env=, models=, header=, name=, style=, chat-path=, models-path=, stream=false, retries=0 — all optional and in any order.

For one endpoint at a time, or full control:

# guided setup
python worker/client.py provider wizard

# a preset, autodetecting its whole catalogue (prompts for the key)
python worker/client.py provider add deepseek --preset deepseek --autodetect

# local Ollama — no key needed
python worker/client.py provider add local --preset ollama

# ANY url, with custom headers and hand-picked tags
python worker/client.py provider add mygw \
    --url https://gw.example.com/v1 \
    --header 'X-Org: acme' --header 'X-Trace: on' \
    --map fast=llama-3.1-8b --map smart=llama-3.1-70b

# keep the key out of the config file
python worker/client.py provider add openai --preset openai --key-env OPENAI_API_KEY

# a gateway with the key in a query param and no model-listing endpoint
python worker/client.py provider add odd \
    --url https://odd.example.com/api/v9 --chat-path /complete \
    --models-path - --auth query --auth-param access_token \
    --map odd-fast=their-model-name

Useful flags for provider add / provider edit:

flag	meaning
--url	provider base URL
--style	openai | ollama | anthropic
--map TAG=MODEL	advertise TAG, routed to MODEL (repeatable)
--autodetect	advertise every model the endpoint reports
--include / --exclude	glob filters over autodetected model ids
--header 'Name: value'	extra header, repeatable; {api_key} is substituted
--key / --key-env / --key-stdin	how to supply the API key
--auth	auto | bearer | x-api-key | query | none
--chat-path / --models-path	override endpoint paths (- = no listing)
--extra-body JSON	merged into every request body
--no-qualify	don't advertise <provider>/<model> aliases
--timeout	per-request timeout (default 300s)
--stream / --no-stream	stream from the endpoint (default on) vs wait for the whole reply
--retries N	extra attempts on transient upstream failures (5xx/524/timeouts; default 2)

Slow providers / gateway timeouts ("524 A timeout occurred")

Some upstreams — especially Cloudflare-fronted proxies like justwoker.icu — kill long, idle connections after ~100s. The worker avoids this by streaming from the endpoint by default (tokens keep flowing, so the connection stays alive no matter how long generation takes), and it automatically falls back to plain requests if an endpoint rejects stream=true. Transient failures (5xx, 524, connection resets, timeouts) are retried a couple of times with a short backoff before being reported, so a flaky gateway gets a second chance instead of failing the request outright.

Those tokens are also forwarded live: while generating, the worker reports progress to the coordinator (POST /api/progress/{id}), which relays it to whoever is waiting on /api/generate with stream: true. Workers and coordinators that predate this still interoperate — an old worker simply reports no progress, and the stream falls back to keepalives plus one final line with the whole answer.

Inspecting and running

python worker/client.py provider list          # what's configured
python worker/client.py provider models mygw   # ask the endpoint what it hosts
python worker/client.py provider test          # connectivity + auth for all
python worker/client.py provider test mygw --prompt 'hi'   # real generation
python worker/client.py tags                   # tag -> provider:model table

python worker/client.py run                    # donate, using every provider
python worker/client.py run --only deepseek,mygw
python worker/client.py run --concurrency 4

Providers live in one JSON file (chmod 600, it holds API keys):

python worker/client.py config-path
# override with PLATINUM_WORKER_CONFIG=/path/to/providers.json

Hot reload: edit providers while the worker runs. Adding, removing or remapping tags is picked up within a few seconds and re-announced to the coordinator — no restart, no dropped jobs. Disable with --no-reload.

Tag precedence

1. explicit --map tags (your intent always wins)
2. autodetected model ids
3. <provider>/<model> aliases, always unambiguous

If two providers expose the same bare name, the earlier one wins it and the other stays reachable through its qualified alias.

Legacy usage (still supported)

python worker/client.py                                   # local Ollama
python worker/client.py --backend llamacpp --endpoint http://localhost:8080
python worker/client.py --backend openai \
    --endpoint https://api.example.com \
    --models glm-4.7-flash,glm-4.5-flash

These bypass the config file entirely and behave exactly as before, including --ollama-host / --ollama-port / --api-key / EDITORAI_WORKER_API_KEY.

Your privacy: the worker never sends your IP address to the coordinator. The network is pull-based (your machine polls for work; the coordinator never connects back), so your address is never needed, transmitted, or stored. API keys never leave your machine either — only the tag names do.

---

🛠 For Coordinator Operators

Fallback chains map a requested tag onto an ordered list of models to try. The built-in table is unchanged; to define chains over your own tags, point PLATINUM_CHAINS at a JSON file (re-read automatically when it changes):

{
  "best":     [["mygw/llama-3.1-70b", 2], ["deepseek-chat", 1]],
  "cheapest": ["deepseek-chat", "fast"]
}

[model, tries] pairs, bare strings (1 try), or {"model": "x", "tries": 2} objects all work. Tags without an entry get a single attempt, so any tag a worker advertises is servable with no config.

---

📊 Monitoring

# See workers and stats (coordinator)
curl http://sn-1.vltgg.net:21800/api/status | python -m json.tool

# Check health (coordinator on :21800, public proxy on :21801)
curl http://sn-1.vltgg.net:21800/health
curl http://sn-1.vltgg.net:21801/health