a WRAITH product
Any model. Same evidence.
Candor AI sits in front of every provider your policy approves: frontier labs, crypto-native and confidential-compute networks, or models you run yourself. Each call gets the same policy check, DLP, record and on-chain anchor, whoever served it.
Every modality, one audit trail
Each endpoint gets the same treatment: identity, role-based model access, DLP, budgets, a write-ahead record sealed with post-quantum encryption, and an on-chain anchor. Prices are set per modality in the policy, so cost records are exact.
Text, code and agents
Chat completions and Anthropic Messages, streaming, tool use, and agent payments. /v1/chat/completions · /v1/messages
Embeddings
Vectors for search and retrieval. Inputs are scanned; the record keeps the vector count, dimensions and a hash. /v1/embeddings
Image generation
Prompts are scanned before the provider sees them. Each image is hashed into the record and priced per image, with a per-request cap. /v1/images/generations
Text to speech
Voice output for assistants and accessibility. Input text is scanned; the audio is hashed and kept in the sealed record. /v1/audio/speech
Transcription & translation
Meeting and call audio to text. Transcripts are scanned and redacted before they reach the caller. Priced per minute. /v1/audio/transcriptions
Video roadmap
Video generation has no common API across providers yet. We'll add it per provider as enterprise demand and standards settle.
Audio uploads can't be scanned for sensitive data, so they're refused unless your policy explicitly allows audio uploads. When the policy redacts output, transcripts are returned only as plain text or JSON text, never as segments or subtitles that could split sensitive data. Endpoints work with any provider that offers them through an OpenAI-compatible API.
Bring your own provider accounts
You connect your own accounts with each provider, so your contracts, data-processing terms and payment method stay between you and them. The policy pins each provider's endpoint and lists its attestations. A HIPAA policy, for example, will only route to providers marked with a BAA.
| Provider | Type | API | Paying the provider | Data handling (per the provider) |
|---|---|---|---|---|
| Anthropic | Frontier lab | Anthropic native | Card or invoice | Enterprise terms, zero-retention options |
| OpenAI / Azure OpenAI | Frontier lab | OpenAI | Card or invoice | Enterprise terms, regional deployments on Azure |
| Venice | Crypto-native, private | OpenAI-compatible | Card, crypto credits, staked DIEM, or pay-per-call USDC (x402) | "Private" models are not stored, according to Venice; TEE and end-to-end encrypted modes are available |
| NEAR AI Cloud | Confidential compute | OpenAI-compatible | Pay as you go, digital assets, staking credits | Intel TDX and NVIDIA GPU TEEs with signed responses; DPA available; no training on customer data |
| Phala Private AI | Confidential compute | OpenAI-compatible | Prepaid balance | TEE inference with per-response attestation receipts; zero-data-retention models |
| Chutes | Decentralized (Bittensor) | OpenAI-compatible | Card, crypto or TAO, pay per token | Terms say API content is not logged or stored; TEE models available |
| AkashML · io.net | Decentralized compute | OpenAI-compatible | Per provider | Check each provider's terms |
| OpenRouter | Aggregator | OpenAI and Anthropic formats | Card or crypto | Zero-data-retention setting per account |
| Your own models | Self-hosted | OpenAI-compatible (vLLM, Ollama, TGI, LM Studio) | Your infrastructure, or verified GPUs on WRAITHGRID | Never leaves your network |
Provider details are summarised from each provider's public documentation as of October 2026 and can change. Privacy claims are the provider's own. Candor AI records what was sent and returned; it cannot audit what happens inside a provider. Where a provider offers hardware attestation (TEE), keep its attestation with the record.
Choose per workload, in policy.
Different roles can use different routes. For example, engineering can use frontier models, legal can use confidential-compute models only, and agents can use your self-hosted fleet. Each record states the route it took.
Your own models
Open-weight models on your hardware or private cloud. Maximum control, no third party in the data path.
Your provider accounts
Frontier labs and crypto-native networks under your own contracts and payment methods, crypto included.
Candor Relay in development
Prepay in USDC on Monad and route to supported providers through Farnsworth, with per-call receipts anchored on chain. Opt-in only: the relay puts Farnsworth in the data path, which the self-hosted modes never do.
Need a provider that isn't listed?
Any endpoint that speaks the OpenAI or Anthropic API works today. Tell us what you use and we'll add a preset for it.