gmtrenchor-mcp

GMTRENCHOR measures what is normally invisible on Arc and Hyperliquid, keeps it, and publishes its own scoreboard, failures included.

Start with track_record. It is free, it states the lift of this scanner against pools it never called, and it shows the score bands that do not separate and the theses our research engine has refuted. A signal service that will not publish its own failures is asking to be taken on trust.

The order to work in

1. What launched — token_microstructure for the first hour: buyers, concentration, the price at ten, thirty and sixty minutes. Arc's public nodes prune the swap tape within days, so this window cannot be fetched from anywhere at any price once it has passed. Ours was written down while it existed.
2. Can it be sold again — that is a live test with real money and it lives at a different service, honeycheck. Nothing here repeats it; the token answer points at it and carries the verdict that was already published in the pool's first hour.
3. Who deployed it, who holds it — deployer_history.
4. What the market is doing — market_context, and carry_persistence if you are sizing a funding trade rather than reading one.
5. What happened to calls like it — open_calls, against track_record.

Tools

| Tool | What it answers | | | --- | --- | --- | | track_record | What this scanner called and what became of it, with the lift, the bands that fail, and what the lab has refuted. | free | | open_calls | Every call open on Arc now, with what was known when it was opened. | 3¢ | | token_microstructure | One token's first hour, frozen before the chain forgot it. | 5¢ | | deployer_history | What one address has launched and what became of it. | 5¢ | | market_context | Hyperliquid now against the same hour a day ago. | 1¢ | | carry_persistence | How long funding spreads have actually lasted. | 5¢ |

No tool here gives advice. Each answers a question about the past with the denominator beside it. A rate over eleven observations and a rate over eleven thousand read the same in prose, so every answer states which it is.

Use it

{ "mcpServers": { "gmtrenchor": { "command": "npx", "args": ["-y", "gmtrenchor-mcp"] } } }

With nothing else set the free tool works and the paid ones say what they need.

Paying per call

| Variable | What | | --- | --- | | PAY_WALLET_KEY | The agent's own key. Read from its own environment, used only to sign the payment authorisation locally, never sent anywhere. | | PAY_CHAIN | Where that wallet's USDC is: arc (default), base, arbitrum or hyperEvm. The payment need not be on the chain the data is about. | | PAY_MAX_USDC_PER_CALL | The most one call may cost. Default 0.05, which admits every tool here. The owner sets this. A price above it is refused before anything is signed. |

Payment is USDC over x402 through Circle Gateway, gas-free for the buyer.

You are never charged for an answer that was not produced. The service verifies the payment, produces the answer, and only then settles.

What has and has not been exercised

Audited on 18 September 2026 with Circle's own client and a real signature from an unfunded wallet: on every route and on all four chains Circle verified the signature, the answer was produced, and settlement was refused for one reason only, insufficient_balance, with the answer withheld.

The data sets are bought once at a URL rather than through a tool, because a tool that hands a model nineteen dollars of gzip is a bill and not an answer: every closed call with its outcome at https://gmtrenchor.memeogatchi.workers.dev/x402/calls/closed, and one frozen day of every Arc launch at https://gmtrenchor.memeogatchi.workers.dev/x402/panel?day=YYYY-MM-DD, listed free at https://gmtrenchor.memeogatchi.workers.dev/panel/days.

Read provenance in the closed-call set before you parse it. Arc rows are complete, because we read that chain ourselves block by block. Rows from the other chain we scan carry the claim, its time, the score and the reasons but not the prices: those observations came from a third party whose terms forbid redistributing or deriving from them, so the columns are null and each row names which. The rows are kept rather than dropped — a record showing only the chain where the scanner looks good would not be a record.