Apps for humans.
Market data for AI agents.

The same liquidation tape, sold per call in USDC on Algorand mainnet. Four endpoints: volume for one symbol split by exchange and by whether longs or shorts were flushed, a cascade score for whether that volume landed in one burst or spread evenly across the window, the full symbol universe so a caller can tell coverage from silence, and per-exchange reporting integrity — two of the three venues throttle to one update per second and undercount exactly when it matters. Every response says what it is: measured, absent when the market genuinely had none, or unmeasured when our own lookup broke. Unmeasured returns 503 and is never billed, because charging for a zero we invented would make every other number worthless. Discoverable in the x402 Bazaar. No signup, no API key.

A pre-flight execution guard that sits between an agent's decision and the chain. It came out of auditing seven live data collectors and finding eight real bugs of one family: a broken lookup recorded as a finding. A rate limit written down as "no exit exists". A health check reporting healthy on a database it could not read. So abstain splits evidence three ways — measured, absent when the world genuinely said none, and unmeasured when our own pipe broke — and refuses to let anything be decided on unmeasured. Every path writes a receipt, including every refusal. Deterministic idempotency keys mean a crashed retry replays instead of double-spending. 160 tests, zero dependencies, MIT.

Anyone can put a dataset on chain. Nobody checks the bytes. corpus makes each dataset epoch an ERC-721 vault and puts a machine appraiser between the data and the buyer. The appraiser reads the actual file, runs eight checks covering schema, row count, null rate, types, time ordering, label ordering and outliers, scores it, signs the report and posts the verdict on chain. A refusal is not a low score, it is a hard gate: the purchase reverts and the vault cannot be sold even by its owner. Two vaults are live on X Layer mainnet holding production data. The launch outcomes vault carries 74,802 labels checked against their own observation window close, zero violations, so its no-lookahead property is machine verified rather than asserted. Delivery is hash gated end to end: the gateway finds a file by hashing candidates against the content hash the contract stores, so it cannot serve bytes the vault never committed to. 21 contract tests, 13 gateway tests, MIT.

Models rarely die loudly. A column turns to VARCHAR, a table stops landing, a pipeline gets rewired under the feature store, and nothing throws an error while the predictions go quietly wrong. fray walks DataHub lineage backwards from every production deployment through the model, its features and every table upstream, and freezes that closure as the model's contract. Nobody configures anything. It rescans, diffs, and ranks every break by blast radius, meaning which live endpoints inherit it, so on call reads fraud detector prod endpoint at risk instead of table has nulls. Findings are written back into the graph as assertions with failure run events, heartbeats on the clean tables, tags and a handoff brief, so the next scan starts from what the last one learned. 11 tests, Apache 2.0.

FTSO publishes the price. haircut publishes whether that price is reachable. FAssets mints FXRP against collateral on Flare, and every collateral ratio you can read on chain is computed at the oracle price. Price the FLR leg at what SparkDEX would actually pay and it recovers 0.09% of its marked value — agents whose pool ratio reads 209–650% land at 0.20–0.61% once repriced, against a 150% minimum. This is not a solvency claim: the USDT0 leg covers FXRP three times over on its own, the Core Vault holds 7.4M XRP against 1.77M FXRP outstanding, and every agent is fully backed. The exposure sits with collateral pool providers. Readings land in an append-only registry on Flare mainnet that stores refusals as first-class entries, because a gap nobody recorded is indistinguishable from a period nobody looked.

A lending market marks collateral with an oracle. A liquidator only acts if the liquidation bonus beats the cost of getting out. overhang measures the difference across every reserve on Solana's tokenized-equity lending markets, quoting real sell-side exits at size and bisecting for the largest clip that still clears. Running every 15 minutes since 11 August: 2,158 sweeps and 25,896 measurements with zero gaps, each one carrying a size-matched control quote so a failed pipe is never published as an absent market. On the xStocks board 82.6 percent of marked collateral cannot be realised, dollar-weighted. The sharper finding is on the other market: Superstate's Opening Bell tokens are transfer-restricted to allowlisted wallets while Kamino's liquidation instruction is permissionless, and the router refuses to quote them at all — TOKEN_NOT_TRADABLE on 2,167 of 2,167 measurements, $24.0M of collateral whose liquidator market cannot exist. A terminal verdict requires six consecutive agreeing observations, so one bad quote never becomes a finding. Curators and oracle vendors are paid by the protocols whose parameters they set. This is the independent read, and it is sold per query in USDC. 115 tests, MIT.

A forecaster that publishes its own scorecard before it knows the answer. It reads a 28-day tape of 1.4M perp liquidations across 799 symbols and returns the probability that a symbol liquidates more in the next 15 minutes than its own 90th-percentile window — 0.15 Brier skill against climatology on 243,000 held-out questions. Every forecast is written to a public log before the window it describes and settled afterwards from the exchange's own feed; the database refuses to let a prediction be written late, edited, or settled twice. The scoring half is built on proper scoring rules, so base-rate camping, cherry-picking, copying and confidence inflation each score worse than answering honestly. A probability edge, not a trading edge. 99 tests, zero dependencies, MIT.

Agent memory that can be cross-examined, built on HydraDB. Every answer carries the claims it rests on and the traversal that found them, and every refusal carries what was searched and why nothing qualified. Scored on all 500 LongMemEval questions: 0.3511 accuracy against a no-memory baseline of 0.000, graded by a disclosed judge stronger than the answerer, with zero questions left ungraded. When it answers it is right 70.5 percent of the time, and 28 of the 30 unanswerable questions were correctly refused. The finding is that the failure mode is over-refusal rather than hallucination: 133 questions were refused with the gold session sitting in the evidence, and every one of them is clickable in the published register. Refusing with the answer present and refusing with nothing retrieved are counted separately, because they need opposite fixes. 334 tests, zero dependencies, MIT.

Evidentiary memory for agents, on CockroachDB. An agent that remembers everything it was told remembers the lies too. alibi keeps provenance attached to every stored fact so a claim can be traced back to who said it and when, and quarantines what cannot be vouched for rather than letting it into the answer path. Built on CockroachDB's managed MCP server and distributed vector indexing, with the cluster and object storage on AWS. 127 tests, MIT.

Depth-adjusted collateral limits for Blend on Stellar. A lending market sets its limits from an oracle price, but the number that decides whether a liquidation actually clears is what the collateral can be sold for at size. sounding measures the gap and publishes the depth the oracle price does not carry, so a limit reflects the exit that exists rather than the print. Architecture published in the repo rather than described. Soroban, MIT.

The same per-call model on a third chain family. Three priced endpoints — ticker, spread, volatility — plus a free discovery route, metered in USDC over x402 on Aptos. Payment challenges arrive base64 encoded in a response header rather than the body, and the facilitator sponsors gas, so a caller spends USDC and nothing else. No signup, no API key. Testnet, with mainnet held deliberately as future work rather than shipped early.

Five collectors publish a dated finding every day under CC BY 4.0, with permalinks, JSON-LD and a machine feed. One measures what lending collateral can actually be sold for: two Superstate reserves on Kamino have never routed at any size across more than 1,100 consecutive sweeps, while Kamino's liquidation instruction stays permissionless.

31,228 paired observations of tokenized US equities against the stock they track — 11 symbols on Raydium, 19–29 July 2026, on a 300-second cadence at 99.1% coverage. Median deviation is 24 bps while the underlying is trading and 52–58 bps when it is not, because the reference price freezes at the closing bell and the token does not. On MSTRx and CRCLx the gap runs 7.5× wider — the two most crypto-correlated names in the basket, repricing news no closing print can hold. Raw ticks and a daily aggregate on Hugging Face, chart on Dune, CC BY 4.0.

Drop AgentFeed into any elizaOS agent — one npm install and it gets the AgentFeed catalog: liquidation tape, squeeze scores, funding, orderbooks, whale flow, Solana on-chain and DeFi, paying per call for exactly what it uses.

Trust ratings for tokenized real-world assets. Five transparent pillars — peg integrity, backing, issuer, liquidity, redemption — scored from live on-chain data, with every score showing its reasons.

Sharp-money detection on pre-match odds. It watches the line move against the money and flags the moves that come from informed volume rather than public sentiment — the ones books adjust for quietly. Built for the TxODDS Trading Tools track, where it took 2nd place and 4,000 USDT. The interesting engineering problem was proving a detector works when it has correctly fired zero times: no alerts turned out to mean thin fixture coverage, not broken code, and the difference was only provable by driving a synthetic tick through the live path end to end.

Live World Cup alerts for people who cannot watch the match. Goals, cards and state changes pushed to Telegram the moment they happen, built for the Consumer & Fan Experiences track. The detail that mattered was starting the watcher before kickoff rather than at it — a bot that wakes up when the match begins has already missed the first goal.

The starter kit I wish had existed before shipping eight apps to the Solana dApp Store. Seed Vault wiring, wallet session handling, the publishing pipeline and the parts of the store submission that are only obvious after you have failed them once. Extracted from production apps rather than written as a demo, so the defaults are the ones that survived review.

On-chain AI assistant, native to the Seeker. Ask it anything — crypto questions come back with live SOL/BTC prices, funding and sentiment. Price alerts ping you even with the app closed.