YieldSignal
About
Real-time risk-weighted USDC lending APY across Aave, Compound, Morpho, Moonwell, Euler and Fluid on Base, paid per call via x402
Details
- Author
- stakemate369
- Categories
- Finance
Jump to
Setup
Install YieldSignal in your MCP client (Claude Desktop, Cursor, Windsurf, and others).
Repository: https://github.com/stakemate369/yieldsignal
Follow the installation instructions in the repository README, then restart your MCP client.
Risk-weighted yield signals for autonomous agents, paid per call via thex402protocol — no API key, no signup. Every product route is paid:$0.10signal,$0.25analytics (durability/capacity/sensitivity/exposure),$0.50decision,$1.00persistence. There is no free tier and no bypass parameter.
- ETH liquid staking(Ethereum mainnet) acrossLido, Rocket Pool, Coinbase Wrapped Staked ETH, Frax Ether and Binance Staked ETH
- USDC and WETH lending(Base) acrossAave, Compound, Morpho, Moonwell, Euler and Fluid
Seven products: the rawsignal(what pays best right now), thedecision(given where your money already sits, is moving it worth the cost — MOVE/HOLD with expected net gain and break-even in days), thedurabilityreport (how much of that APY survives if incentives stop), thecapacityreport (can you actually withdraw your size from that market), thesensitivityreport (how close the market is to the kink where borrow rates explode), theexposurereport (how much of your portfolio sits behind the same risk, no matter how many venues it is spread across), and thepersistencereport (how long a call here actually holds before the leader flips — measured from this service's own on-chain attestations, which is why no competing feed can produce it).
GET https://yieldsignal.vercel.app/signal/eth-staking-yield GET https://yieldsignal.vercel.app/signal/usdc-base-yield GET https://yieldsignal.vercel.app/signal/weth-base-yield GET https://yieldsignal.vercel.app/decision/eth-staking-yield?position=lido&amountUsd=25000&horizonDays=30 GET https://yieldsignal.vercel.app/durability/weth-base-yield GET https://yieldsignal.vercel.app/capacity/usdc-base-yield?amountUsd=200000 GET https://yieldsignal.vercel.app/sensitivity/usdc-base-yield GET https://yieldsignal.vercel.app/exposure/usdc-base-yield?positions=aave:200000,morpho:150000 GET https://yieldsignal.vercel.app/persistence/usdc-base-yield
Bare/signaland/decision(no asset) redirect to the ETH staking route — the asset with the strongest verified track record. Short forms like/signal/usdcredirect to the canonical path.
Don't take our word for it:GET /accuracy.jsonisfreeand returns the per-asset within-tolerance hit-rate and average regret in bps, computed 1:1 from the public on-chainEAStrack record. As of writing, the staking signal holds up far better than the USDC lending one (Base USDC churns too fast for a call to stay true) — which is exactly why the flagship is staking. The numbers are recomputed from chain data, not asserted here, so check them yourself before paying.
Call it without payment first and you'll get a402 Payment Requiredwith the exact price/asset/network to pay. Any x402-compatible client can complete the payment automatically — for example with@x402/fetch:
import { wrapFetchWithPayment } from "@x402/fetch"; import { CdpX402Client } from "@coinbase/cdp-sdk/x402"; // or any other x402 client/signer const client = new CdpX402Client(); // needs CDP_API_KEY_ID/SECRET/WALLET_SECRET + a funded wallet const fetchWithPayment = wrapFetchWithPayment(fetch, client); const res = await fetchWithPayment("https://yieldsignal.vercel.app/signal/usdc-base-yield"); console.log(await res.json()); // { bestProtocol: "compound", gapBps: 57, rates: [...], asOf: "..." }
Also available as paid MCP tools athttps://yieldsignal.vercel.app/mcp—get_yield_signalandget_yield_decision, both with an optionalasset("ETH_STAKING","USDC"or"WETH"; defaults to"USDC"for backwards compatibility with already-published integrations). Most autonomous-agent frameworks discover/call tools via MCP rather than hand-rolled x402 HTTP clients. Uses the official@x402/mcppackage; payment is gated per tool call (tools/list/initializestay free, only the tool call requires payment). There's also anelizaOS pluginin the official registry.
Every rate is tagged withwhere it came from—onchain/api(Aave, Compound, Morpho — read directly from the protocol) ordefillama(Moonwell, Euler, Fluid — via the DefiLlama yields API). No estimated or fabricated numbers: a source that fails or returns invalid data is omitted from the response, never guessed at.
Every rate is also compared onone explicit basis(apyBasis: "supply-apy-total-incl-rewards"): base interestplusincentives. This matters because the raw sources disagree on what "APY" means — an on-chainliquidityRate/getSupplyRateis base-only, while DefiLlama's aggregate and Morpho'snetApyalready include reward tokens. Ranking them side by side without reconciling that compares different things. Each rate itemizesapyBaseBps/apyRewardBpswhere the source separates them, andrewardBasissays how the incentive component was obtained (reported,inferred,included-not-itemized,unavailable). Any protocol whose incentive could not be established at all is listed inincompleteRewardData— its APY is a floor, not a verdict.
{ "asset": "USDC", "bestProtocol": "compound", "gapBps": 150, "rates": [ { "protocol": "compound", "apyBps": 601, "apyBaseBps": 601, "apyRewardBps": 0, "rewardBasis": "reported", "tvlUsd": 8446455, "tvlBasis": "total-supplied", "weightedApyBps": 595, "source": "onchain", "asOf": "2026-07-30T..." }, { "protocol": "moonwell", "apyBps": 434, "apyBaseBps": 403, "apyRewardBps": 31, "rewardBasis": "reported", "tvlUsd": 2643239, "tvlBasis": "aggregator-reported", "weightedApyBps": 382, "source": "defillama", "asOf": "..." } ], "omittedProtocols": ["euler"], "coverage": { "read": 5, "expected": 6 }, "apyBasis": "supply-apy-total-incl-rewards", "incompleteRewardData": [], "asOf": "2026-07-30T..." }
Depth, and whether the yield is a campaign
Each rate also carriestvlUsdwith atvlBasissaying what that number measures —total-suppliedwhen it comes from the protocol's own books (AavetotalAToken, ComettotalSupply, MorphototalAssetsUsd),aggregator-reportedwhen it comes from DefiLlama, which reportsavailable liquidityfor lending markets. They are not the same quantity and the field says which one you got. Depth matters because the top of a thin market is not an executable rate for size: at the time of writing, the WETH signal leads with a $716k market whose yield is 59% incentive campaign.
The decision routes act on both:/decision/returnsgainDependsOnIncentivesandpositionShareOfDestinationPct, states them in the human-readablereason, and lowers confidence when your position would be a large share of the destination. A real response:
MOVE, confidencemedium— "Moving aave → euler yields +106bps risk-adjusted … Note: the entire gain rests on euler's incentive campaign (172bps of its 291bps) — it disappears if the campaign ends. Note: your $200,000 would be 27.9% of euler's $716,897 market — large enough that entering dilutes the rate you are moving for."
Splits every protocol's APY into base interest vs incentive and reports thepost-incentive floor: what you keep if the reward campaign stops. On a live reading (2026-08-05), WETH on Base led witheulerat 299bps —57.9% of it incentive, floor 126bps — whileaave's 153bps was entirely base. Without incentives the ranking flips. The signal endpoint alone would have pointed you at euler.
Two rules keep this from becoming a false-alarm generator:
- Only itemized sources are decomposed.rewardBasis: "included-not-itemized"or"unavailable"means the floor is unknown, and the protocol is listed inundecomposable— never treated as incentive-free. Absence of a reported incentive is not evidence of absence.
- No ranking claim when the leader is opaque.If the current best protocol can't be decomposed,rankingChangesWithoutIncentivesisnull, notfalse. You still getbestVerifiableFloor— the highest yield provably independent of incentives — which is a fully measured statement either way.
Base lending only (USDC/WETH).Checked live on 2026-08-05, all five liquid-staking sources reportapyReward: nullon DefiLlama — 0 of 5 decomposable on every call — so an ETH staking durability route would charge for "cannot establish". Staking yield being incentive-free is plausible, but a missingapyRewarddoesn't prove it, and that's the same inference this service refuses to make forfluid/morphoon the lending side.
Deliberately not a date forecast.The obvious source for "when does this campaign end" is Merkl's API (earliestCampaignEnd). Checked live on 2026-08-05 againstapi.merkl.xyz/v4/opportunities?chainId=8453&status=LIVE: Aave, Compound, Euler and Fluid havezeroLENDcampaigns there, and Morpho/Moonwell appear only as curated MetaMorpho vaults that aren't the markets this service reads.status=PASTreturns null timestamps, so renewal history isn't measurable either. Covering 2 of 6 protocols by guessing at vault matches — with most campaigns being weekly and auto-renewing — would produce a confident-looking number that is wrong, which is the opposite of what this service sells.
A lending market at 99% utilization pays beautifully and won't let you withdraw; the high rateisthe symptom.tvlUsdcan't tell the two apart — $100M supplied with $99M borrowed and $100M supplied with $10M borrowed are the same number.
Utilization and free liquidity come from the protocol's own books, in thesame call that already fetched the rate(Aave: thegetReserveDatatuple; Compound:getUtilization()+totalSupply()), so this costs zero extra RPC. Pass?amountUsd=and each protocol reports whether that size can exit now (canExitNow), the coverage ratio, and what share of the market it would be.
Morpho and the DefiLlama-sourced protocols don't publish borrowed-vs-supplied, so they're markedmeasured: falseand areneverreturned asbestProtocolExecutable. Unmeasured is not the same as liquid. USD figures are USDC-only — converting a WETH balance would need a price oracle inside a paid response path; WETH still gets utilization, which is unitless.
How close is the repricing? —/sensitivity/
A lending rate is not a number, it is a function of utilization with akink. Below it the rate creeps; above it, it explodes. Read live on 2026-08-05:
utilization kink headroom borrow APY at 90% → 93% compound USDC 89.82% 90% 0.18pp 4.08% → 15.95% aave USDC 86.16% 90% 3.84pp 4.60% → 7.79%
Compound's USDC market waseighteen hundredths of a pointfrom the knee. The signal endpoint alone shows a healthy 3.29% supply APY and says nothing about that.
The kink and both slopes are read from each protocol's own interest rate contract — for Compound the reconstructed curve is checked againstgetBorrowRate(u), the contract's own pure function, and matched it wei-for-wei at 50/85/90/93/99% utilization; for Aave the shape is checked against the ceiling the contract itself reports (base + slope1 + slope2 == getMaxVariableBorrowRate). If either check fails the protocol is dropped rather than served with a number derived from a formula that changed.
This is also the first route that speaks to theborrowerrather than the lender.
Aave and Compound only.Morpho'sAdaptiveCurveIRMfails on three independent counts, checked live: its curve constants aren't exposed on-chain (CURVE_STEEPNESS,TARGET_UTILIZATION,ADJUSTMENT_SPEEDandINITIAL_RATE_AT_TARGETare allinternal constant, so using them would mean hardcoding numbers copied from a repo); whatisreadable,rateAtTarget(id), isstate that drifts over time, not a static curve, so "the rate at 95%" has no fixed answer there; and this service reads avault(Gauntlet USDC Prime, $429M across 5 Blue markets with different collateral) whose curator reallocates at will. Moonwell, Euler and Fluid come from DefiLlama with no curve access at all. All of them are reported asmeasured: false— never as stable.
One risk in three wrappers —/exposure/
GET /exposure/usdc-base-yield?positions=aave:200000,compound:50000,morpho:150000
The market is full ofevent detectors— depeg alerts, hack alerts, liquidation alerts — and most of them are free. None of them answer the question that actually costs money:am I exposed, and through what path?
In the Stream Finance collapse, only1 of ~320 MetaMorpho vaultsheld the broken asset directly ($700k of bad debt), yet $93M of loss became$285M of contagionbecause the exposure arrived indirectly. The event was public within hours; what hurt people was not knowing they were two hops from it.
A real reading on 2026-08-06 for the portfolio above:
$400k across 3 venues — $200k attributable collateral cbBTC 81.0% via compound+morpho ← shared curator 0x9e33fa… 75.0% via morpho collateral WETH 13.4% via compound+morpho ← shared parameter kink=9000 100.0% via aave+compound ← shared unattributed: aave $200k
Three venues, and 81% of what can be attributed sits behind one collateral. Thekinkline says something else: Aave and Compound both reprice at exactly 90% utilization, so splitting capital between them buys no protection at all against a utilization shock.
Attribution differs by protocol, because the risk topologies differ, and thebasisfield says which one you got:
- Morpho(isolated-market) — each Blue market is isolated with a single collateral, so the attribution is exact.
- Compound(collateral-basket) — one base asset against a defined basket, weighted by what isactually posted(totalsCollateral× the Comet's own price feed). Measured: cbBTC 43.1%, WETH 37.4%, tBTC 7.8%, cbETH 6.9%, wstETH 4.8%.
- Aave— reportedunattributed, and not for lack of effort: a v3 supplier is exposed to the entire pool's collateral set. Attributing it to one asset would be false, and splitting it across all of them would imply diversification that does not exist.
Percentages are ofattributablecapital, not of the total — over the total they would shrink and read as low concentration when the truth is low visibility.coveragereports both.
This measuresstructural shared exposure— the factual claim "these positions depend on the same thing" — not statistical correlation. Saying how much they move together would need a price history this service does not have, and would be a weaker claim dressed as a stronger one.
Recursive-collateral detection is deliberately absent.Checked live across the 77 live Morpho markets on Base: zero cycles, and zero assets that are both collateral and loan asset. Worse, the recursion that killed Stream was not in the lending graph at all — it was in theissuanceof the synthetic (xUSD backed by positions funded with the borrowed USDC). A cycle detector would have reported "all clear" throughout the collapse. False safety is worse than a false alarm.
How long does the answer hold? —/persistence/
Every other endpoint here — and every competing yield feed — answerswhat pays best now. None answershow long that stays true, which is the number that decides whether moving capital is worth the gas.
This one answers it from this service's own hourly attestations on Base. Over 24 days and 439 attestations, the three assets are not the same product at all:
Half of those 167 USDC switches are around trip between the same two protocols(compound <-> fluid). Following that signal literally costs more in gas than the edge is worth — and until this endpoint existed, nothing in the catalogue said so.
It also settles, empirically, an assumption almost everyone makes:a bigger lead does not last longer.Spearman-0.04across 177 completed leadership spells, with the0-24bpsand>=300bpsbands both sitting at a 2h median. Gap size is not a usable proxy for confidence, and this is the only place that number is published — including when it is unflattering to the service publishing it.
The/decision/routes consume this internally: expected gain is projected over the horizon you asked fororthe measured lead duration, whichever is shorter. A 30-day projection on a 2h edge is exactly how a MOVE that only pays gas gets recommended.
Honest limits, stated in the payload rather than in a footnote: the observation interval is 1h, so leads shorter than that are invisible; assets whose lead has not yet changed report afloor with a censoring flag, never a median; and every figure travels with its sample size, withnullwherever the sample is too small. Every input is a public attestation UID — recompute the whole report frombase.easscan.organd you get the same numbers without trusting this server.
GET /accuracy.json(free) carries two independent measures, both derived from the public on-chain attestations:
- score— directional: was the flagged protocol still the leader (or within 25bps)when scored against the market right now.
- windowedScore— each attestation judged overits own validity window, i.e. until the next attestation for that asset replaced it.medianWindowHoursis the practical answer to "how often should I re-check?" and it differs sharply per market: on the record as of 2026-07-30,13h for ETH liquid staking and WETH lending, 1h for USDC lending on Base. USDC rotates fast; that is a property of the market, not a defect the endpoint hides.
The track record now covers more than the signal
The whole argument of this service isverifiable, not promised— and until 2026-08-06 the EAS record coveredonly the signal. The four analytics products were sold leaning on credibility the signal had built, contributed nothing back to it, and could not be scored.
SENSITIVITY_SCHEMAis the first to change that, and sensitivity was chosen to go first for a specific reason: it is the only product whose record turns into anempirical questionlater. With utilization and kink stamped on-chain at each reading, the history eventually answers"markets that sat within half a point of the kink — how often did they cross, and how fast?"*— and the headroom number stops being merely descriptive.
Onlymeasuredentries can be attested (encoding a non-measured one throws), and only marketswithin 5 percentage points of the kinkare written: attesting everything on every read has no cost ceiling, and a market parked at 60% utilization produces no information a month later would not.
Gated byEAS_SENSITIVITY_SCHEMA_UID. Empty means the trigger simply does not run — the same switch-by-omission the v2 signal migration used. Registering it costs real gas and requires a typedCONFIRMvianpm run register-schema.
Full write-up with every on-chain artifact, the exact command to verify each one, and an honest list of what the stack doesnotsolve:docs/verifiable-agent-trust.md.
Two independent ways to check a response wasn't tampered with or fabricated, neither requiring you to trust this server's uptime at the moment you check:
- Signed responses— every REST/MCP response is signed (EIP-712 typed data) by the samepayToaddress the 402 payment requirement names. The struct (asset,bestProtocol,weightedApyBps,gapBps,asOf,contentHash) mirrors the on-chain EAS schema below, plus acontentHash(keccak256of the exact response body) binding it to the full response. REST:X-Signal-Signature/X-Signal-Signer/X-Signal-Eip712-Payloadheaders. MCP: a sibling content block. Verify withviem'sverifyTypedData— or just callgetSignalVerified()from theyieldsignal-clientpackage, which does both checks for you.
- On-chain attestations(EAS, Base mainnet) — periodic, public, permanent records of "at time T, protocol X paid Y bps, Z ahead of the runner-up," independently checkable oneasscan.orgwithout trusting this server at all. Same attester address as the signed responses above. Published automatically whenever the signal changes materially (best protocol flips, or the gap moves ≥25bps) or gets stale (>12h since the last one) — seesrc/attestation/autoAttest.tsandPOST /internal/auto-attest(cron-triggered, not on every paid call — that would have no cost ceiling).npm run atteststill exists for manual, on-demand publishing. Full history atGET /track-record(or/track-record.json).
- Agent discovery(ERC-8004, Base mainnet) —GET /agent-card.jsonis the registration file; agentId59272already minted onIdentityRegistry. Any REAL buyer (a wallet other than the service's own — the contract blocks owner/operator self-feedback) can leave verifiable feedback onReputationRegistryby runningnpm run give-feedback— addresses/ABI insrc/attestation/erc8004.ts.
Sibling project to YieldPilot (a personal Aave/Morpho/Compound rebalancer), butfully separate: own CDP credentials, own receiver wallet, no shared runtime code. SeeCLAUDE.mdandSECURITY.mdfor the full technical writeup and threat model.
- src/expressApp.ts— the Express app + x402 payment gate (createX402Server,@x402/express), reused by both the local dev server (src/server.ts) and the Vercel serverless entrypoint (api/index.ts). Registers one payment-protected route per asset (RESOURCE_PATHS) plus an unpaid/healthliveness check.
- src/signal/— the pure, deterministic comparison logic (no I/O, fully unit tested).
- src/market-data/— the two-layer data sourcing (direct reads + DefiLlama), parametrized by asset (AssetId,USDC/WETH) viaconfig/networks.ts'sBASE_ASSETS.
- src/wallet/walletLock.ts— pins the receiver wallet address (viaEXPECTED_WALLET_ADDRESSin production, since serverless has no persistent disk) so a CDP credential rotation is caught loudly instead of silently redirecting payments.
- src/mcp.ts— theget_yield_signalMCP tool (optionalassetparam), gated per-call with@x402/mcp'screatePaymentWrapper(not the whole-route Express middleware, which would paywalltools/list/initializetoo).
- src/market-data/cache.ts— 30s TTL on every rate reader (direct + DefiLlama), one cache instance per asset, so a burst of concurrent paid calls doesn't hammer public RPC/API endpoints.
- src/notify/paymentLog.ts— logs payer/tx/network/amount for every settled payment (onAfterSettle, both the REST and MCP payment servers) and alerts on a payment from a wallet that isn't the owner's (a real sale).
- src/usage/— durable funnel counters (usageStore.ts+usageMiddleware.ts), readable atGET /usage.json(auth required). Counts 402s served, paid attempts, settlements, failures and 404s per route/asset. The sale-side stages (paid_attempt/served/settled/failed) are exempt from the daily event budget — crawler noise used to exhaust it before midday and erase exactly the rows that answer "does this sell?". Exists because platform runtime logs aren't reachable from outside, so the only auditable signal used to be on-chain revenue — which can't distinguish "nobody arrives" from "they arrive and don't pay". Backend is any Redis with a REST API; credential discovery matches by env-varsuffix, so a store connected through the Vercel dashboard (which prefixes the vars it injects) is picked up with no code change. With no store configured it degrades to per-instance in-memory counters and reports"durable": false.
- src/cli/withdraw.ts— sweeps accumulated USDC to the owner's personal wallet, manualCONFIRMrequired, never automatic.
- src/wallet/signerAccount.ts— resolves the same receiver wallet withsignMessage/signTypedData/sendTransactionexposed (createX402Serveronly exposes the address); used for response signing and EAS attestation.
- src/attestation/— EAS schema definition and calldata encoding (pure, unit tested);publishAttestation.ts(shared tx-sending logic),queryAttestations.ts(EASScan GraphQL client + decoder),autoAttest.ts(pure decision logic + orchestration for the automatic trigger),trackRecord.ts(then-vs-now comparison),erc8004.ts(ERC-8004 registry addresses/ABI).
- src/cli/registerSchema.ts/src/cli/attestSignal.ts/src/cli/registerAgent.ts/src/cli/giveFeedback.ts— one-time schema registration, manual per-attestation publishing, one-time ERC-8004 identity mint, and buyer-side reputation feedback, sameCONFIRMpattern aswithdraw.ts.giveFeedback.tsis the one script that deliberately resolves a DIFFERENT wallet thanwallet/signerAccount.ts— the contract rejects feedback from the service's own address.
npm install npm test # automated tests (market-data readers, signal logic, retry, wallet lock, usage funnel) npm run signal # live signal, real data, zero credentials needed (npm run signal -- ETH_STAKING / -- WETH for the other assets) npm run dev # local x402 server (reads X402_ENVIRONMENT from .env) npm run test:paid # spins up a test buyer wallet, funds it via the CDP faucet, pays for real (testnet only, REST endpoint) npm run withdraw # sweep accumulated USDC — asks for typed "CONFIRM" npm run register-schema # one-time EAS schema registration (mainnet, real gas) — asks for typed "CONFIRM" npm run attest # publish one on-chain attestation of the current signal (mainnet, real gas) — asks for typed "CONFIRM" npm run register-agent # one-time ERC-8004 identity mint (mainnet, real gas) — asks for typed "CONFIRM" npm run give-feedback # a REAL buyer leaves feedback on ReputationRegistry (mainnet, real gas) — asks for typed "CONFIRM". Fails with "Self-feedback not allowed" if run from the service's own wallet.
See.env.examplefor the required variables. Generate your own dedicated CDP project/credentials atportal.cdp.coinbase.com— never reuse another project's.
Why Spark, Seamless and Silo aren't in the protocol list
Checked againstyields.llama.fi/poolson 2026-07-16, filteringchain=Base+symbol=USDC: Spark only has a USDS pool (not USDC) on Base; Seamless and Silo have no indexed Base pool at all right now. Rather than fabricate a number, these three are left out until a real USDC market exists for them on Base — see the comment insrc/market-data/types.ts.
Checked live againstyields.llama.fi/poolson 2026-07-17: there's no canonical "WBTC" market on Base, only Coinbase'scbBTC(a different asset), and where it does have a market its supply APY sits at ~0-0.2% across all six protocols — a signal too flat to be worth selling ("which is best" barely matters when everyone's tied near zero). WETH, by contrast, has real, distinct yield across all six protocols (0.66%-3.7%), so it's the asset that got added instead.
Bridge Town is an MCP-native, git-versioned financial modeling platform for FP&A teams and finance leaders. AI agents use Bridge Town tools to create projects, write Python model files, run models in isolated cloud sandboxes, query data, write outputs to Google Sheets, create dashboards, branch scenarios, and collaborate with teammates.
The Capital.com MCP Server lets your AI assistant talk to your trading account directly. Market data, position checks, trade previews – all in plain language, without leaving your AI tool.
Coinrule Agentic Trading MCP enables investors to create, backtest, execute, and manage trading agents through natural language across stocks, crypto and ETFs
Invest with Claude and other AI assistants
Australian Consumer Data Right Product Data
Remote MCP server for historical crypto & prediction-market data: search ~500K instruments, live market stats (OHLC, turnover, spreads, depth, slippage) and tick-data purchase. Keyless for catalog & stats; optional OAuth for account tools. Endpoint: https://cryptostruct.com/mcp
Cross-border debt collection from your AI assistant: check cases, get pricing, submit new cases.
Read-only MCP server for your Evibe investment portfolio + live market data (holdings, performance, dividends, benchmarks, screeners). Works with Claude & ChatGPT.
Financial and quantitative modeling engine for AI agents. Typed, named, deterministic.
Sign in to leave a review
Use Google, GitHub, or an email account so ratings stay tied to real people.
No reviews posted yet.
