OpenNodes

@opennodes/cli

The OpenNodes Protocol (ONP) client: discover verified AI inference nodes, pre-price multi-step chains with enforceable estimates, invoke with pinned prices and signed receipts — and bring all of it into the tools you already use.

• `onp mcp` — a stdio MCP server for Cursor, VS Code, Claude Desktop, Kiro and any MCP host. Tools: recommend (best node + model for a task, priced and explained; the prompt never leaves your machine — only extracted features are sent), search_offerings, get_offering, get_node (vet a provider: operator, tier history, availability), estimate. Local models and memory: recommend shows, for models you run yourself (the Ollama library), the memory each size needs (4-bit estimate: 8b ~6 GB, 70b ~46 GB) and, given policy.local_memory_gb, leaves out what cannot fit and picks the largest size that does. This server fills that value from the machine (75% of RAM) when the agent does not pass it, and says so in local_memory.
• `onp gateway` — a local OpenAI-compatible server over discovered nodes. Point any OpenAI-SDK app at it. Virtual models onp/auto, onp/auto-cheap, onp/auto-fast, onp/auto-quality, onp/auto-private, onp/auto-eu (EU/EEA operators only), onp/auto-ca or any onp/auto- route each request by the advisor; aliases and query-defined virtual models are supported; every call is receipted. Failover: auto models try up to 5 ranked nodes in turn; a node that hasn't started answering within a budget built from its measured latency and the request size (15 s – 3 min) is skipped; and an OpenRouter-style models: ["a", "b"] list in the request adds further models to try, in order (it is never forwarded to nodes). Auto-routing only picks what you can actually call: free registered nodes, plus the listings of any host you hold a key for — passed with --key host=KEY or picked up from OPENROUTER_API_KEY, HF_TOKEN, OPENAI_API_KEY, MISTRAL_API_KEY, GROQ_API_KEY, TOGETHER_API_KEY, DEEPINFRA_API_KEY, FIREWORKS_API_KEY (each key is sent only to its own host; --no-env-keys turns detection off). Trust floor: registered nodes join the pool from the community tier up, and imported listings (always unverified) score lowest on trust; --min-tier verified refuses anything below verified for every request, named models included (403 policy-refused), and --min-tier unverified also admits registered nodes that haven't passed their checks yet. Passive reputation: every call is remembered per offering — outcome and timing only, never prompts or answers — in ~/.opennodes/reputation.json (last 50 calls, 30 days; --no-reputation to keep it in memory). With 3+ calls it fills success rate and time to first token where the registry has measured nothing (imported listings), and an offering with 3 failures in a row is benched for 15 minutes; registry measurements from blind probes always win. Advice says "your calls: 19 of 20 ok, first token 820 ms". See /gateway/reputation. Embeddings: POST /v1/embeddings (OpenAI shape, single or batched input). onp/auto uses the embedding models already pulled into your local Ollama first (--ollama http://127.0.0.1:11434 by default, --no-ollama to skip): free, and the documents never leave the machine; then registered nodes and hosted embedders on providers you hold a key for, ranked by size, popularity, price and trust, with failover. onp/auto-private is local only; onp/auto-eu / onp/auto- filter hosted ones; an offering id or a bare local name (nomic-embed-text) picks one directly. Cost is logged from reported tokens × the listed price; local calls cost $0.
• `onp check` — your own quick verification of a listing the registry hasn't verified (imported catalogs, registered nodes awaiting their checks), with your key and a spend cap (default $0.02, priced before each request; listings without a published price are estimated at $10/$30 per MTok): the model is in /models, it answers, token counts are reported and plausible, time to first token and throughput, a fact surviving a long prompt (--context 8192, or full for 90% of the claimed window), and declared JSON mode / tool calls. onp check / , or --url /v1 --model for any OpenAI-compatible endpoint. Results are saved in ~/.opennodes/checks.json for a week, and the gateway uses them: what failed to answer is skipped by onp/auto, your measured speed fills in where the registry has none, failed JSON/tool support is dropped and a failed context size caps the window. --json for machine-readable output, --no-save to keep it out of routing.
• `onp search | estimate | invoke | card` — the CLI.

npx @opennodes/cli mcp --print-config      # paste into your MCP host's config

npx @opennodes/cli gateway --port 4141 # then set OPENAI_BASE_URL=http://127.0.0.1:4141/v1 npx @opennodes/cli search --modality text --tier verified --sort tps

Registry: ONP_REGISTRY — defaults to the hosted registry at https://registry.opennodes.io (9,000+ offerings); set it to http://127.0.0.1:4300 to use your own (npx @opennodes/registry).

Part of OpenNodes · specification · Apache-2.0.