AI Agent / Web3·Solana · EVM
Galadriel
Verifiable LLM inference layer for onchain agents. The Sentience SDK is a drop-in OpenAI-compatible client that runs inference inside a TEE (teeML), produces a Proof of Sentience attestation per call, and posts those attestations onchain on Solana — letting agents prove which model produced which output without rewriting prompts or tools.
- 01verifiable LLM inference for agents
- 02OpenAI-compatible drop-in with attestation
- 03onchain proof-of-output for agents
- 04fully onchain unruggable AI agents
- 05TEE-based verifiable Claude/GPT calls
- pip install sentience
- pnpm add @galadriel/sentience openai
| Variable | Scope | Description |
|---|---|---|
| GALADRIEL_API_KEY | Server | API key from dashboard.galadriel.com used to authenticate verifiable inference calls. |
| GALADRIEL_BASE_URL | Server | Endpoint base URL, typically https://api.galadriel.com/v1 — used as the OpenAI client `baseURL`. |
Use Galadriel's Sentience SDK as a drop-in OpenAI client that produces verifiable inference. In TypeScript: `import OpenAI from 'openai'; const client = new OpenAI({ apiKey: process.env.GALADRIEL_API_KEY, baseURL: process.env.GALADRIEL_BASE_URL });` then `const res = await client.chat.completions.create({ model: 'gpt-4o', messages });` — the response includes a `proof` (TEE attestation) and a `tx_hash` for the onchain attestation posted on Solana. For verification call `client.verify(proof)` (or fetch the Solana tx and check the program account) before treating the output as canonical. Supported models include OpenAI and Anthropic frontier models plus open fine-tunes; pick the model name from the supported list and treat the proof as part of the response payload, not metadata.
- ⚑Attestations are posted to Solana — every verifiable call costs a small SOL fee; budget for it and keep the relayer wallet funded.
- ⚑TEE attestations prove the call ran the declared model in the declared enclave, not that the model itself is safe/unbiased — apply normal output guards.
- ⚑Model availability inside the TEE changes — frontier closed models (GPT, Claude) are routed through licensed providers and can be paused without notice; pin a model and monitor the supported list.
- ⚑Pricing per-token includes a verifiability premium over raw OpenAI/Anthropic rates; compare per-call cost when proof isn't required.
- ⚑Region availability is limited to where Galadriel has TEE-capable hardware (mostly US/EU); APAC latency can be noticeably higher.
- ⚑Image / multimodal inference support lags text — confirm the model supports vision before sending image content blocks.
- ⚑On-chain congestion on Solana can delay attestation finality from seconds to a minute; do not block UX on `tx_hash` confirmation, verify async.
- ⚑OpenAI compatibility is high but not perfect — function calling, streaming, and tool_choice behavior can drift per model; snapshot your prompt suite when upgrading models.