/docs/why-run-a-node
Why run an IICP node?
If you have a GPU running Ollama, vLLM, or any local inference backend, you already have everything you need. IICP lets you put that capacity on a shared mesh — every task builds your reputation, and CIP worker mode earns you credits you can spend routing your own tasks.
Join a real mesh
Your node appears in the live directory the moment it registers. Other proxies discover it by intent and route tasks to it automatically — no config on their side.
Earn credits via CIP
Enable CIP worker mode (allow_remote_inference = true) to earn IICP credits for every sub-task you serve. Credits let you route your own tasks across the mesh — a built-in incentive loop, no money involved.
No vendor lock-in
IICP is a protocol, not a platform. Your node registers with an open directory and speaks a documented JSON format. You can switch backends, move regions, or run your own directory.
What actually happens when you run a node
- 1
Your node registers with the directory
On startup, iicp-node sends your endpoint, supported intents (e.g. urn:iicp:intent:llm:chat:v1), and available models to the directory at iicp.network. Registration takes under a second.
- 2
You appear in discover results
A client can discover your node by intent. The directory considers several routing signals, including load and reputation. New nodes start at 0.50; the deployed reputation model currently mixes task outcomes and latency.
- 3
Tasks arrive directly at your node
The directory handles discovery only. Task payloads go from the client to your node — the directory never sees them. Your backend URL and local model server stay private, but your node can read the tasks it executes, so apply the retention and logging policy you advertise.
- 4
Task outcomes provide evidence; eligible CIP tasks earn credits
Completed-task counters are advisory. The deployed reputation score can rise, remain unchanged, or fall depending on reported outcome and latency, and it is not a fraud verdict. Eligible CIP work can additionally produce a signed credit receipt that the node submits to the directory.
Cooperative Inference (CIP) — what it means for you
CIP is Phase 5 of the protocol. When you opt in (allow_remote_inference = true), other nodes can sub-contract tasks to you as a worker — you receive a subtask, run inference, and submit a signed receipt. The coordinator node credits you for the tokens you produced.
You can also act as a consumer: route your own heavy tasks across multiple CIP workers, get results aggregated (best-of-N, majority vote, or map-reduce), and pay with credits you earned from serving others.
Credit rate: roughly 1 credit per 1,000 tokens produced (varies by node pricing). A 2,000-token task earns ~2 credits. Credits have no monetary value — they are mesh-internal capacity tokens used to buy inference from other nodes.
What you need
- ✓A machine with any LLM backend — Ollama, vLLM, llama.cpp, LM Studio, MeshLLM, or another OpenAI-compatible server. You can also run a Claude-backed node natively with the Anthropic backend (--backend-type anthropic) and an API key — no local GPU required.
- ✓A reachable endpoint — direct routing is preferred, but current clients can use Cloudflare Tunnel or configured relay fallback when normal IPv4/IPv6 reachability is unavailable or unverified.
- ✓~20 minutes for first-time setup. IICP is in Beta; the SDKs are publicly available and interested operators can join.
No router access or static IP? Free tunnel options →
| Setup | RAM | Models you can serve |
|---|---|---|
| CPU only (any x86_64 / Apple Silicon) | 8 GB | 3B–7B parameter models (qwen2.5:0.5b, phi3:mini) |
| CPU + modest GPU (6–8 GB VRAM) | 16 GB | 7B–13B models quantized (int8/int4) |
| GPU workstation (24 GB+ VRAM) | 32 GB+ | 34B–70B models fp16 or larger quantized |
Apple M-series Macs use unified memory — 16 GB is enough for 13B models at reasonable speed.
The mesh right now
checking mesh…The directory at iicp.network is the reference Genesis Seed — the bootstrap point for Phase 1 and Phase 5 of the protocol. Full probe results are visible at /stats and /nodes. The network is in early bootstrap — this is the right time to join if you want founding-node reputation before the mesh scales.
pip, npm, or cargo package and run iicp-node serve. The setup guide also provides an installer you can inspect first. Questions or to contribute: [email protected].