Skip to main content
Pegana is Brazilian stablecoin infrastructure on Solana: peg state, depeg simulation, loop exposure, webhook keys — 43 operations across 39 paths, in production, used by real money. It is also a good API. Run through our agent-readiness scorecard it grades A (100/100) — first-call-correct 100, hygiene 100, agent-friendliness 100, security 100 (re-measured against the live spec on 2026-08-06; it scored 99 in July and improved). That matters for everything below. A specification tells you what a call looks like. It cannot tell you whether it will work — and that is structural, not a defect anyone can fix by writing better docs. What follows is not a list of things Pegana got wrong. It is a list of things no specification can carry, at any version, at any quality — which is exactly why a good API is the honest place to show them. Concretely: GET /v1/assets/{symbol}/state with symbol = "string" is precisely what the spec permits. Live, it is a 404. The spec is right and the call still fails.
Every number on this page is measured from a real run against the live API. The scorecard, the compression figure, and the cross-API join were all produced by the commands in quickstart — nothing is estimated.

​
The question is the interface

Before. An agent that wants a stablecoin’s peg state has to know Pegana exists, work out which of 43 operations answers that question, find that the path is /v1/assets/{mint}/state, and handle authentication. Each of those is a separate chance to guess wrong, and every wrong guess costs a round trip and a slice of the context window. After. The question is the interface. Asked in plain Portuguese —
it resolved to GET /v1/assets/USDC/state and came back:
First call, live, no API key — the operation is public, and Gecko knows which ones are because auth is resolved per operation at call time rather than assumed for the whole surface. The agent never sees a credential either way.

​
The join lives between two APIs

Before. “Is the stablecoin in this headline still pegged, and what is it trading at?” is one question that spans three surfaces — a news feed, Pegana, and a price source. The fact that joins them is a token mint. No specification can declare a relationship to a specification it has never heard of, so the join has to live somewhere else. After. The mint join across Pegana ↔ Birdeye ↔ Jupiter is declared once, confirmed against real responses, and then the composed chain plans correctly — offline, keyless, $0, before a single live call is made. That is the difference between a list of tools and a graph: the agent doesn’t rediscover the connection on every task, and it doesn’t invent one that isn’t there.

​
The context bill

Before. Pointing an agent at a 43-operation API means the raw surface competes with the actual task for the context window. The agent spends budget reading about the API instead of using it. After. 79.3% compression on this spec — 113,072 B of raw OpenAPI down to 23,382 B of agent surface. Measured 2026-08-06, not estimated. The number is the least interesting part. Shrinking a spec is easy if you’re willing to drop things; the constraint that makes it worth anything is that first-call-correctness is held while you do it. Compression that starts producing wrong calls has not saved you tokens, it has moved the cost somewhere harder to see.

​
What actually changed

Not the model. Not the agent. Not Pegana’s API — which grades A. What the call was built from.

Try the live surface

Or connect the hosted one