Should my agent call the gateway or the upstream API directly?
Use the Caddy gateway URL. x402 enforcement, payment headers, and settlement happen at the gateway edge. Discovery/OpenAPI should also be fetched through the gateway in production.
Use the Caddy gateway URL. x402 enforcement, payment headers, and settlement happen at the gateway edge. Discovery/OpenAPI should also be fetched through the gateway in production.
That is expected. The gateway advertises payment requirements through
PAYMENT-REQUIRED so your client can sign and retry with
PAYMENT-SIGNATURE.
No. Amarok never stores mnemonics, never signs payments, and never submits orders.
Execution quotes return unsignedTxnsBase64 with
meta.executionSubmitted: false. Your agent signs and submits via its own algod.
No. Payment signing stays client-side in your agent or wallet integration. This site only documents the flow and links to machine-readable contracts.
Use the asset id published in discovery/OpenAPI x402 metadata for the target network. On
Algorand mainnet the default documented id is 31566704.
No. This site is onboarding documentation only. Research and execution-quote data remains API-only behind x402 payment gates. Free probes include discovery, OpenAPI, and execution shapes.
Production remote MCP: https://amarok-mcp.compx.io/mcp (Streamable HTTP over
POST; GET /mcp returns 405 because SSE is not offered). Verify
with https://amarok-mcp.compx.io/.well-known/mcp. For local development, run
@amarok/mcp over stdio — see MCP docs.
No. The remote MCP is walletless. Your agent or client wallet signs USDC payments and passes
paymentSignature when retrying paid MCP tools. Never put mnemonics in MCP
server configuration.
amarok_get_scan return timedOut or capped?
MCP clients typically abort around ~60s. Production scan caps
(ALPHA_SCAN_ORDERBOOK_LIMIT=12, ALPHA_MAX_MARKETS_PER_SCAN=0,
ALPHA_SCAN_TIMEOUT_MS=45000) keep a cold orderbook pass inside that budget.
capped means live markets were truncated; timedOut means the
45s orderbook budget elapsed and remaining books were skipped. Ranking tools share the
same pass. Raise the caps only for HTTP clients that can wait longer.
No. /v1/polymarket/* and /v1/kalshi/* are scan-only
families. Each has its own in-process TTL cache (not Alpha scanCache; not Redis/KV).
Explicitly deferred: execution, WebSockets, Perps, cross-venue matching, quotes/bundle SKUs,
watchlist/history, and GET /quotes. Light GETs can use
AMAROK-SESSION; neither venue /scan is session-covered. See
venues.
Use MCP when your agent host supports MCP tools (Cursor, Claude Desktop). Use direct HTTP when building a custom client against the Caddy gateway. Both hit the same public gateway for paid data; MCP is a typed pass-through wrapper.
Run the protocol API and Caddy gateway from the monorepo, then point
AMAROK_API_URL (stdio MCP) or website env vars at your local gateway.
Cursor / Claude Desktop configs for the hosted Worker are on the
quickstart. Paid tools spend mainnet USDC; use
npm run test -w @amarok/mcp (in-process stub) when you need a no-spend dry-run.