FAQ

General#

What is Elyra? A self-hosted router that sends your OpenAI-compatible requests to the best peer by price, speed, reputation or privacy, and pays providers per token on-chain. See the README.

Is this production ready? No. It is early: testnet, plus a capped mainnet beta (escrow and registry deployed on Robinhood mainnet, deposits capped at $20 per wallet by default). The escrow is unaudited, TEE verification is proven only against a simulated Google, and peer discovery is URLs, gossip and the on-chain registry. See STATUS and MAINNET.

Is the market on the landing real? No. "600+ models / 150+ providers" is a launch target. A fresh node shows a simulated demo market, always labelled SIM; use --no-demo to remove it. Real models are the ones your own sources and live peers provide.

How do you count models and vendors? You can measure any node or /models endpoint yourself with npm run catalog:stats (node tools/catalog-stats.mjs --data .elyra or node tools/catalog-stats.mjs --openrouter --rules). Our deduplication and classification rules (tools/catalog-stats.mjs) are explicit:

1. Simulated and offline sources are excluded: the built-in 29-model / 59-peer demo catalog (SIM) and unreachable sources (OFFLINE) never count as live supply and are reported separately.
2. Meta-routers and moving ~ aliases are excluded: aggregator auto-routers (openrouter/auto, openrouter/free, router/auto, or negative-priced routers like typesafe/jev-router and nvidia/switchyard) and moving alias pointers (~vendor/*-latest) route to other models rather than being distinct model weights.
3. Billing and throughput tiers collapse; parameter sizes stay distinct: :free, :batch, :extended, :nitro, :floor, :online, :thinking and :latest suffixes on vendor/model:tier IDs collapse to one model, while Ollama parameter-size tags (llama3.2:1b vs llama3.2:3b) stay separate and only quantisation suffixes (-q4_K_M, -fp16, -bf16) are stripped.
4. Instruct/chat suffixes, cross-source open-weights names, and redundant date pins collapse: trailing -instruct, -chat and -it tags are stripped; plain open-weights names (llama3.2:1b on Ollama and meta-llama/llama-3.2-1b-instruct on OpenRouter) map to the same canonical key (meta/llama-3.2-1b) while distills (deepseek-r1:8b) and fine-tunes stay separate; date-pinned snapshots (openai/gpt-4o-2024-11-20) collapse into their undated base model (openai/gpt-4o) when that base model is also listed.
5. Vendor prefixes are normalized: leading ~ markers are stripped and multi-prefix makers (meta + meta-llama, bytedance + bytedance-seed, mistralai + mistral, x-ai + xai, moonshotai + moonshot, z-ai + zhipu) count as one vendor; unslashed local names infer their maker from the family prefix.
6. Hosted APIs are separated from P2P and local capacity: today the live node (app.elyrainference.com) serves 1 live local model (llama3.2:1b, 1 vendor; 5 unique models across 3 vendors when the remote GPU tunnel and remembered peer are online). Pointing a node at OpenRouter's public /api/v1/models list (460 raw entries across 69 raw prefixes) yields 339 unique canonical models from 57 vendors after excluding 9 meta-routers and 19 moving ~ aliases and collapsing 82 :batch/:free tier duplicates, 1 -chat duplicate, 10 redundant date snapshots, and 2 duplicate vendor prefixes — all counted under hosted (HOSTED · CLEAR), never peer-to-peer and never private/TEE.

Can I try it without installing? Yes: app.elyrainference.com is a real node with a real model. Ask it something; the page shows which machine answered. It is rate-limited and uses a small model (PANEL).

Can I pay peer sellers from the browser without running a node? Yes, when a public node enables browser session payments (ELYRA_PUBLIC_PAY=1). Your browser generates a local session key, you fund that session address from your wallet (depositFor(sessionAddress), capped at $20 USDG per wallet by default while the contract is unaudited), and your browser verifies each seller receipt against the price you accepted before signing the cumulative EIP-712 voucher. The public node only relays messages and never holds your funds or session key, and browser withdrawals (deposit - owed) leave any signed-but-unredeemed seller vouchers covered. Today ELYRA_PUBLIC_PAY is verified end-to-end on a local chain with a simulated wallet and kept off on app.elyrainference.com while there is 1 live seller; see Wallet and USDG.

What is the difference between a hosted source (like OpenRouter) and a peer-to-peer seller?

• Peer-to-peer seller (LIVE): another Elyra node reached directly over HTTPS or through a NAT relay (/r/<id>). Prompts travel in a sealed X25519 + AES-256-GCM envelope, usage is recorded in an ed25519-signed receipt, payment settles per token on-chain in USDG straight to the seller through ElyraEscrow, and the seller can qualify for TEE · VERIFIED once its Confidential Space image digest is in your trust list.
• Hosted source (HOSTED · CLEAR): an ordinary OpenAI-compatible API key (OpenRouter, OpenAI, Groq, Together) that you attach to your own router under Connect. Hosted models are ranked by their published per-model prices (never treated as free), billed outside the chain by that company under your account, never re-shared with peers or listed on the public market, and never counted as peer-to-peer or private/TEE. Can the website control my node? No. Your panel is http://localhost:8545/app/, an SSH tunnel for a remote node, or, if you opt in, your own hostname behind a login: REMOTE-ADMIN.

What does it cost? Running your own model through your own node is free (price 0). Buying from a peer costs the peer's price per MTok, paid in USDG; on testnet that is free mock money.

Are my prompts private? What are the limits of TEE today?

• Sealed transport (X25519 + AES-256-GCM) hides prompts and completions from network observers and relays, not from the machine running the model.
• LOCAL (your own machine or LAN) is private because your prompt never leaves your hardware.
• REMOTE (a model on another machine reached over an SSH tunnel or VPN, such as a rented GPU) and HOSTED · CLEAR (OpenRouter, OpenAI) are not private: the remote host or API provider can read prompts.
• CLAIMED is a seller's self-declared flag and is ignored by teeOnly and auto:private.
• TEE · VERIFIED requires Google Confidential Space (AMD SEV-SNP or Intel TDX) where Google's signed attestation token binds {id, pub, box, wallet} and matches a container image_digest in your pinned trust list. Current limits: the verifier (router/tee.mjs) is built and tested against simulated Google tokens, but has not yet been run inside a real Google Confidential Space VM (milestone M1 in TEE); ordinary VPSs cannot produce hardware attestations; and without Cloud KMS (M3) a TEE node's identity and wallet reset on reboot. The router itself never logs prompts or completions.

Is it OpenAI-compatible? GET /v1/models and POST /v1/chat/completions (JSON and SSE). Not embeddings, images, audio or the Responses API.

Install and run#

npm start says Node is too old. Needs Node 20 or newer (INSTALL).

The service fails immediately with 203/EXEC. ExecStart points to a node the elyra user cannot read (for example under /root). Install Node to /opt/node.

It refuses to start: passphrase. The data dir is encrypted and ELYRA_PASSPHRASE is missing or wrong. There is no recovery without it.

npm test fails with "node did not start" or random failures. The machine is busy. Boot can take 5 s or more under load; tests wait 30 s now. Run them on an idle machine.

Port 8545 is in use. --port 8546 (and pass a different --data dir if running two nodes).

Ollama and models#

/v1/models has no Ollama model. Check curl 127.0.0.1:11434/v1/models. ELYRA_UPSTREAM_URL must end in /v1. Or Connect → Detect local.

First request takes about a minute. Ollama is loading the model into RAM. Warm requests take about a second on a 1B model.

Out of memory. Pick a smaller model; see the sizing table in OLLAMA.

Routing and payments#

I connected Ollama but auto still answered with canned text. Fixed: a live listing now always beats the simulated market. If you still see [elyra demo engine ...], update the node, or check x-elyra-simulated and that your source appears under Models as LIVE.

"No route". Your policy filtered everything out: teeOnly with no private listing, a maxPrice below every listing, or the requested model has no listing. Loosen the policy or add a source.

A peer shows but is never used. Both sides must use the same escrow and settlement mode (sim vs chain) or there is no route. Also check the protocol version.

A paid request returns 402/503. A provider refuses when your wallet's unpaid balance plus the request's worst case exceeds its credit limit (creditUsd, default $0.50), when your wallet's on-chain escrow deposit exceeds the seller's per-wallet deposit cap (402 deposit_cap_exceeded, default $20 USDG while the contract is unaudited), or when it cannot read the chain (503, fail-closed). Deposit within the cap, or wait for the provider to redeem.

My deposit looks lower than I sent. Check the explorer. Real USDG's supply controller can burn from any address; the mock token cannot. The escrow cannot protect against that (SECURITY).

Mint does nothing / errors. Mint works only for the mock test token on testnet.

Withdraw says locked. Withdrawals need request, then wait 1 day, and must be executed within 7 days.

Website#

Unknown URLs on the site return 404. Intended: the bundle ships a 404.html.

Contributing and security#

See CONTRIBUTING and SECURITY. Report vulnerabilities privately through GitHub Security Advisories.

Running it for other people#

Why is generation so slow on my server? Language models are limited by memory bandwidth, and a small shared CPU server has very little (a typical result is about 1 token per second for a 1B model). Elyra only forwards the request. Use a faster machine or a GPU machine reached over an SSH tunnel: Use your own model.

Is a model on another machine private? No. If its operator can run software on it, they can read your prompts. Elyra labels such a source REMOTE and never counts it as private routing. Only a local model, or a verified TEE peer, is private.

What happens when the model machine goes down? The node notices within about 30 seconds, marks the source OFFLINE, stops offering its models and routes around it. A public try page shows "offline" instead of failing. When the machine answers again it comes back on its own.

What should I back up? The data directory (identity, wallet, state) and your ELYRA_PASSPHRASE, in different places. See Run it in production.

My node found a seller but shows no models. Two nodes can only trade when they use the same escrow. Overview says which escrow the seller uses; check it against the one in the docs and click Use this settlement (or set ELYRA_NETWORK and ELYRA_ESCROW before starting). A node with no escrow configured uses a simulated ledger and cannot trade with an on-chain seller.