Skip to content

Verathos Integration

Verathos SN96 is a decentralized verified-inference fallback target for XEL. Chutes SN64 is the immediate decentralized inference default.

When a Living Character has an encrypted persona artifact and a user sends a chat request, the XEL runtime can select Verathos from provider_policy and send an OpenAI-compatible chat completion request.

Flow:

  1. XEL checks profile visibility, password/payment access, spend limits, and runtime provider policy.
  2. Runtime obtains the allowed persona prompt/context through the configured access policy.
  3. Runtime sends an OpenAI-compatible request to Verathos.
  4. Verathos generates the response and returns any available verification/payment metadata.
  5. XEL records usage, payment receipt, provider ID, model ID, and optional proof refs.

Live model target verified from https://api.verathos.ai/v1/models on 2026-07-06:

qwen3.6-27b-abliterated

The model is explicitly selected because XEL needs an adult-capable, less-censored provider option for consenting-adult character conversations. That does not remove XEL policy boundaries: the runtime must still block minors, non-consensual sexual content, coercion, exploitation, illegal sexual content, and privacy abuse before or around the model call.

Other adult-capable/open-weight candidates can be added through provider policy after live availability is verified on the provider:

  • Dolphin Llama 3 / 3.1 variants
  • Llama/Qwen/Mistral abliterated variants
  • Mixtral-style roleplay variants
  • Owner-approved custom fine-tunes

Without an API key, this prints the request payload:

Terminal window
npm run example:verathos

With an API key, when Verathos opens API-key accounts:

Terminal window
VERATHOS_API_KEY=vrt_sk_... npm run example:verathos

The adapter lives at:

runtime/providers/verathos.mjs

Verathos publicly presents an OpenAI-compatible API and the live endpoint currently returns an x402 upto payment challenge without requiring an API key. The observed live challenge uses Base USDC:

FieldLive value
Networkeip155:8453
AssetUSDC
Asset address0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913
Schemeupto

The Verathos web app also reports that API-key accounts and prepaid TAO/USDC deposits are still in preview. Signup currently returns 503 API is in preview mode. Public access coming soon.

Because XEL’s default home chain is Sui, the first deployment needs one of:

  • direct Verathos x402 support for the chosen wallet/network;
  • an XEL/GEN payment facilitator that can authorize from Sui and settle to a supported x402 rail;
  • a character-funded operating balance on the provider-supported USDC rail.

The testnet runtime path can now simulate this settlement without real funds: the provider adapter captures the x402 challenge, runtime/payments/x402.mjs builds the unsigned payment attempt and facilitator request, records a simulated testnet receipt, reconciles it, and retries the OpenAI-compatible request with an X-PAYMENT header.

Remaining environment and secret inputs for a real paid Verathos call:

NameSecretPurpose
VERATHOS_API_KEYOptionalBearer token for the Verathos OpenAI-compatible API once API-key accounts are available. The x402 path can request a payment challenge without it.
VERATHOS_BASE_URLNoOverride for https://api.verathos.ai/v1 if Verathos publishes a different testnet endpoint.
XEL_X402_FACILITATOR_URLNoXEL or third-party facilitator endpoint that can verify/settle x402 payment payloads.
XEL_X402_FACILITATOR_TOKENYesOptional facilitator API credential when the facilitator is not public verifier-only.
XEL_RUNTIME_PAYMENT_WALLET_IDNoWallet descriptor ID from payment-capability.v1, normally the character’s inference_runtime wallet.
XEL_RUNTIME_PAYMENT_SIGNER_REFYesReference to the testnet signer/key material held in Infisical or an HSM/KMS; do not store raw keys in repo files.
SUI_RPC_URLNoSui RPC endpoint used by a real Sui settlement adapter to verify balances or transaction refs.
SUI_NETWORKNoExpected settlement network, for this slice testnet or equivalent non-mainnet configuration.

Do not hard-code Verathos as the only inference provider. It is a fallback/alternate provider target, with Chutes SN64, Targon SN4, Phala, Marlin/Oyster, Akash-hosted open models, and GEN runtime as alternate/fallback provider classes.

Verathos remains a direct decentralized inference provider target even if XEL later adds Talus Nexus as an orchestration layer. Talus should be evaluated as a workflow coordinator around a selected inference provider, not as the model backend itself.

A Talus-coordinated Verathos path could:

  1. request or verify allowed memory retrieval;
  2. call the Verathos OpenAI-compatible inference endpoint through a Nexus tool;
  3. coordinate payment, metering, or gas-budget handling;
  4. return the model response and provider metadata to the XEL runtime;
  5. trigger owner-policy-aware memory write-back; and
  6. emit workflow, payment, and receipt references.

The response metadata should still identify Verathos as the inference provider and the selected model ID as the model. Talus/Nexus metadata should be recorded separately as orchestration, workflow, tool, or receipt metadata.