AI inference, peer to peer. Zero cut in between.
Your computer sends each request straight to someone else's machine and pays it directly on Solana. No server in the middle, and the program takes nothing.
waiting for the first requestReplay of the local rehearsal. Nothing leaves this page.
On your computer, OpenAI-compatible
One signature per answer, on Solana
0%
Anyone. No approval, no delisting
Routing. One word decides the trip.
Point anything that takes a base URL at the router: Cursor, Aider, Continue, Open WebUI, the OpenAI SDK. Then add one word after the model name.
- :cheap
- lowest price for the model
- :fast
- quickest first token this router has seen
- :trusted
- most USDC settled on chain, bonded $OREY
- :private
- verified enclaves only, the request sealed
If a provider stops answering, the next one does. In the rehearsal the cheapest provider was switched off mid-run and the request fell through without the tool noticing.
client = OpenAI( base_url="http://127.0.0.1:7171/v1", api_key="any") client.chat.completions.create( model="llama-3.3-70b:cheap", messages=messages)
- 01Kestrel GPU Co-opeu-north$0.12 in · $0.30 outPicked
- 02Ironbridge Computeus-east$0.20 in · $0.45 outnext in line
- 03Vault Nineeu-west$0.25 in · $0.60 outnext in line
answered by Kestrel GPU Co-op229 ms21+23 tok$0.000010paid
Payments. Sign once, pay per answer.
Your wallet opens a channel with a provider. After every answer the router on your computer signs the new running total, and the provider cashes the latest one on Solana whenever it likes.
- #01deepseek-v3+3131signed
- #02llama-3.3-70b+1546signed
- #03llama-3.3-70b+1561signed
- #04llama-3.3-70b+1576signed
- #05llama-3.3-70b+1591signed
- #06llama-3.3-70b+15106signed
- #07llama-3.3-70b+15121signed
- #08llama-3.3-70b+151362tUj6…9DbBN
redeemed in one transaction 5HUR…jztKUM$0.000136 → Ironbridge Compute
- 01Open
One wallet signature puts two dollars in a channel.
- 02Answer
A signed total after each. A provider is never owed more than one answer.
- 03Redeem
Many channels in one transaction, for a fraction of a cent.
- 04Close
What is left comes straight back. A lost key spends at most that, under your daily cap.
Providers. Every one is someone's own machine.
A GPU running vLLM or Ollama, a llama.cpp box, API capacity you already pay for. List it with one command. No application, no stake required, and nothing in the program can delist you.
Run a providerLight a node. Set your price.
Put a provider node in front of any OpenAI-compatible server. It checks each voucher before it answers and redeems the totals to your wallet.
$ orrery provider register \ --config provider.json --wallet wallet.json $ orrery provider start --config provider.json
Registering puts your endpoint and the hash of your price list on chain. Your wallet stays cold; the node runs on its own key and can never touch what you have earned.
- Redeemed to its wallet
- $0.000136
- Answers served
- 8
- Taken by the program
- $0
| Model | In / M | Out / M |
|---|---|---|
| llama-3.3-70b | $0.20 | $0.45 |
| deepseek-v3 | $0.27 | $1.10 |
Prices are yours. The price list's hash is on chain, so routers ignore any list that doesn't match it.
The program in the middle takes0%
It has no fee field and no third account to send one to. Every unit a provider redeems lands in the provider's account. Every unit it doesn't comes back to you.
Privacy. Sealed when you ask for it.
Add :private and the router only picks providers running inside a hardware enclave. It checks the enclave's live report against the claim the provider posted on chain, then encrypts your prompt to a key that never leaves it. The operator hosts the machine and still can't read the request.
$OREY. Backs the providers you trust.
Launching on Meteora. In the program it has one job: providers bond it behind their listing, and routers count the bond toward reputation.
- Bond
- Posted on chain behind a provider's listing
- Reputation
- Bonded OREY is 10% of the
:trustedscore - Never slashed
- Unbonding returns it after seven days
- Contract
- not launched yet
- 13
- requests
- 3
- providers
- $0.000199
- settled on chain
- $0
- to the program
- 14/14
- checks passed
- not yet
- on mainnet
The program, a router and three provider nodes on a Solana test validator, 2026-11-03 04:41 UTC. Mainnet is read from the chain.
All 14 rehearsal checks
- /v1/models lists the network catalogllama-3.3-70b×3 qwen3-32b×2 deepseek-v3×1
- price route picks the cheapest llamaKestrel GPU Co-op, $0.000010
- streaming works and only Ironbridge has deepseekRequests travel from your computer to a …
- privacy route seals the request to the only enclave
- route header works too
- speed route settles on the provider with the fastest first tokenIronbridge Compute
- when the cheapest provider goes dark the request falls throughIronbridge Compute answered
- router logged every answered request13 trips
- Kestrel GPU Co-op redeemed exactly what the router signed$0.000010 over 1 redeem(s)
- Ironbridge Compute redeemed exactly what the router signed$0.000136 over 1 redeem(s)
- Vault Nine redeemed exactly what the router signed$0.000053 over 1 redeem(s)
- a provider releasing its channel sends the rest straight back to the userrefund $0.999864
- no fee in the program: user + providers + open deposits = 100 USDCuser $97.9998 · providers $0.0001 · deposits $1.9999
- program totals agreesettled $0.000199 · 3 channels
program 7UHe…ytZS · 13 program tests · 12 router tests