Decentralized inference network where workers submit predictions for named Topics (price feeds, forecasts, classifications) and reputers score them. Consumers fetch a single context-aware aggregated inference per Topic — useful as a price oracle, signal feed, or model-of-models backend for trading and DeFi agents.
- 01decentralized price/forecast oracles
- 02ensemble ML inference
- 03trading signal feeds for agents
- 04Topic-based prediction consumption
- 05Cosmos-native AI marketplace
- pip install allora-sdk
- pnpm add @alloralabs/allora-sdk
| Variable | Scope | Description |
|---|---|---|
| ALLORA_API_KEY | Server | API key from developer.allora.network for the off-chain inference endpoint. |
| ALLORA_CHAIN_RPC | Server | Allora chain RPC endpoint, e.g. https://allora-rpc.testnet.allora.network for on-chain consumption. |
Use Allora to consume aggregated inference from a Topic. From TypeScript: `import { AlloraAPIClient, ChainSlug, PriceInferenceTimeframe, PriceInferenceToken } from '@alloralabs/allora-sdk'; const client = new AlloraAPIClient({ chainSlug: ChainSlug.TESTNET, apiKey }); const inf = await client.getPriceInference(PriceInferenceToken.BTC, PriceInferenceTimeframe.EIGHT_HOURS);` returns a `network_inference` plus a `confidence_interval`. From Python use `from allora_sdk.v2 import AlloraAPIClient`. For on-chain consumption, query the Topic's latest inference via the Allora chain query module (`x/emissions/network_inference`) or the consumer contract pattern that pins a Topic id and reads the value as an oracle. Always treat the inference as probabilistic — read both the value and the confidence interval, and degrade gracefully when stale.
- ⚑Topics have skewed update cadences (some tick every block, some every hour) — always read `block_height` / `timestamp` and reject inferences older than your SLA.
- ⚑Topic ids and definitions can be re-parameterized by governance; pin Topic id + `topic_metadata` hash and re-validate on deploy.
- ⚑Confidence intervals from the network model uncertainty; ignoring them and using the point estimate as gospel will leak alpha and amplify tail risk.
- ⚑Worker payouts and reputer scores can churn the underlying inference distribution — short-term inference quality varies until a Topic warms up post-launch.
- ⚑ALLO token volatility plus on-chain consumer fees mean per-call cost in USD swings; budget headroom on app deployments.
- ⚑Region/availability of the off-chain API is centralized to Allora Labs — for true decentralization, consume the Topic via the Allora chain rather than the REST endpoint.
- ⚑Allora is forecast/numerical-first — there is no general image or text-generation Topic; do not treat it as a drop-in for LLM inference.
- ⚑Slashing applies to misbehaving workers/reputers, but consumers see no direct refund — design for occasional bad-tick outliers in your downstream logic.