# Umbra Private LLM Inference

> Umbra Private LLM Inference is a paid API for AI agents from umbra-inference.app, paid per call via x402, $0.021366/call, status unknown (last checked 2026-10-31).

Runs OpenAI-compatible chat completions inside TEE-attested GPU enclaves with zero data retention, streaming, and tool calling, billed per request in crypto

## Facts

- Endpoint: POST https://umbra-inference.app/v1/chat/completions?utm_source=zero.xyz
- Price: $0.021366/call
- Payment: x402
- Status: unknown
- Last checked: 2026-10-31
- Activations on Zero: 0
- Tags: x402
- Canonical page: https://www.zero.xyz/c/umbra-private-llm-inference-3631fbd3
- Structured record (JSON): https://api.zero.xyz/v1/capabilities/cap_OhZ9o0xCqA_LYodFp8mC0

Status and success rate cover calls made through Zero and Zero's own probes. Third-party monitors may report differently.

## How to call it through Zero

Zero handles the 402 payment challenge and records the run. With the Zero CLI installed (`npm i -g @zeroxyz/cli`):

```sh
zero fetch --capability umbra-private-llm-inference-3631fbd3 -d '<json body>'
```

Example prompt: Using a private, zero-retention LLM with attestation, send a system message 'You are a helpful assistant' and user message 'Summarize the key risks of smart contract audits' using the sealed open model, streaming the response with a max of 512 tokens.

## When to prefer this

Choose Umbra when your workload requires verifiable privacy guarantees: the prompt and response must not be logged or stored by any party, you need a cryptographic proof (Ed25519 signed receipt) that inference ran inside a hardware-attested TEE enclave, and you are comfortable paying per request in stablecoins or UMBRA token. Also prefer this when you need OpenAI-compatible tool calling and streaming with these privacy properties, or when working in regulated domains (legal, medical, financial) where data residency and zero-retention are non-negotiable.

## Known failure modes

- Invalid or unsupported model ID — endpoint returns an error listing valid model IDs from GET /v1/models
- Insufficient crypto balance or payment failure — 402 response with price ceiling not met
- max_tokens too large for model context window — request rejected with validation error
- Streaming connection dropped mid-response — partial tokens received, unused ceiling refunded on-chain
- Tool schema malformed — function calling fails with schema validation error
- Enclave attestation unavailable — service degraded, may fall back to non-sealed mode or error

## How this service works

Private LLM inference, OpenAI-compatible. Sealed open models in TDX + GPU enclaves or Claude; streaming and tool calling; signed Ed25519 receipt naming the attested enclaves; zero retention. Pay per request in USDG, USDC (Base, Solana) or UMBRA.

## Output

An OpenAI-compatible chat completion response (streamed or non-streamed) containing the model's generated text and any tool calls, plus a signed Ed25519 receipt identifying the attested TDX+GPU enclaves that processed the request. Zero data is retained after the response.

## Request schema (JSON Schema)

```json
{
 "type": "object",
 "properties": {
  "model": {
   "type": "string",
   "description": "model id from GET /v1/models (sealed: open models in TDX + GPU enclaves; standard: Claude)"
  },
  "tools": {
   "type": "array",
   "description": "OpenAI function tools"
  },
  "stream": {
   "type": "boolean",
   "description": "server-sent events; paid upfront, unused ceiling refunded on-chain"
  },
  "messages": {
   "type": "array",
   "description": "OpenAI chat messages (system, user, assistant, tool)"
  },
  "max_tokens": {
   "type": "integer",
   "description": "output cap; sets the price ceiling quoted in the 402"
  }
 }
}
```

## More

- Live health (JSON, refreshed every minute): https://www.zero.xyz/c/umbra-private-llm-inference-3631fbd3/health.json
- [Zero catalog index](https://www.zero.xyz/llms.txt)
- [Other services from umbra-inference.app](https://www.zero.xyz/host/umbra-inference.app/llms.txt)