TrueTick

Demo

About

The same real-world stock is now tokenized by several issuers, on several chains, in several DEX pools — and they don't agree. NVDA alone exists as NVDAon (Ondo, Ethereum), NVDAx (Backed, on both Ethereum and Solana), NVDA (Robinhood Chain) and NVDAc (Coinbase B20, Base). A per-chain dashboard shows you one of them and tells you nothing about the other four. TrueTick puts all of them in one table, priced against the real-world reference, so you can see which price is real. Right now the reliable venues agree within about 0.5–0.8% of each other — and one NVDA pool prices at $1,153 against a real $218. That pool holds $27k and hasn't traded in 24 hours. TrueTick doesn't hide it: it shows the number and flags it as phantom-liquidity and price-divergence. The agreement between the honest venues is what makes the outlier legible. Two rules run through the whole system. Every price field is `number | null`, never 0 — a zero propagates silently through arithmetic and renders as a real price, which is exactly how two silent-$0 bugs in our own subgraph would have shipped unnoticed. And a bad price is always labelled, never dropped: market-closed, thin-volume, phantom-liquidity, subgraph-syncing, source-error. The headline spread is computed across reliable venues only, because including the broken pool would report a spectacular, meaningless "428% spread". The data is the product, so it's sold rather than given away. The price service sits behind x402: an agent requests it, gets HTTP 402, pays 0.01 HBAR settled through the Blocky402 facilitator on Hedera testnet, and receives the data plus a transaction id it can verify on HashScan. An MCP server lets Claude Desktop be that paying agent — it names a ticker, our server decides to spend, and the wallet key never reaches the model. Everything is live on-chain data. Nothing is mocked.

How it's made

The core is one normalized schema (`TokenizedStockPoint`) with one adapter per venue. Adding a chain or a DEX is a new adapter file plus registry rows — the schema, quality rules, HTTP service and UI are untouched. That isn't a claim, it's the record: adding Solana (non-EVM, base58 addresses, no chain id, a DEX with no swap event) required zero schema changes; the types file gained two optional fields. Four data sources, across two Graph products: • Ethereum (Ondo, Backed) — Uniswap v4 subgraph. v4 exposes no USD price, so it's derivedETH × Bundle.ethPriceUSD, cross-checked against the deepest *traded* pool's own ratio. Using the deepest pool by TVL instead flagged a false 4.97% divergence caused by an untraded memecoin pair. • Robinhood Chain (Uniswap V3) — Substreams. The Graph's networks registry lists "subgraphs": [] for eip155:4663, so no subgraph was possible. • Solana (Raydium CLMM) — Substreams. There's no swap event to decode, so price comes from vault balance deltas: decode the CLMM PoolState for the pool's two vaults, diff pre/post token balances, and skip same-sign pairs (those are liquidity adds, not trades). • Base (Coinbase B20, Aerodrome Slipstream) — a subgraph we forked, fixed and published ourselves, because none existed for that CL factory. So we're a Graph publisher, not just a consumer. Both Substreams adapters compose prebuilt StreamingFast foundational modules driven by query-string params — there's no Rust in the repo. The .spkg files are vendored so the pipeline doesn't depend on registry availability at runtime. Publishing the subgraph surfaced two bugs that would have produced a healthy-looking subgraph returning $0 everywhere: upstream whitelisted USDC at Avalanche's address (eth_getCode proves no contract exists there on Base), and getEthPriceInUSD loaded a Uniswap pool that an Aerodrome-only subgraph never indexes, so ethPriceUSD was 0 and every USD field with it. Payments: x402 v2 with the Blocky402 facilitator on Hedera testnet, HBAR (asset 0.0.0), settled as a native HAPI TransferTransaction — so the receipt is a Hedera tx id, not an EVM hash, and the facilitator pays the network fee as co-signer. The client cross-checks the challenge's advertised feePayer against Blocky402's published signers *before* spending, because pointing at a generic x402 facilitator would still settle, still show real transactions, and silently miss the requirement. Settlement happens after the handler returns and is cancelled on a 4xx/5xx, so a failed request costs nothing. The MCP server exposes six tools to Claude Desktop. The hacky-but-necessary part: Claude Desktop cancels a tool call at 60s and a cold Substreams venue takes 105–120s. Paying first risked settling after the client gave up — charged, nothing shown. So get_deviation probes warmth with a free request first and refuses to pay while cold, and concurrent requests for a ticker are deduped into one fetch (three "parallel" calls were queueing on the 2-session cap and all looking like cold starts). Favourite bug: byAddress() lowercased its input, correct for EVM hex and silently wrong for base58 — 'Xsc9qvGR…' is not 'xsc9qvgr…', so every Solana lookup would have missed and the venue would simply have been absent.