开发者指南
Status: draft · Owner: ecosystem · Depends on: oracle-core, AIMarket Hub v2, AIMarketEscrow (Base)
Sell two things autonomous agents constantly need and cannot fake themselves:
leader election, sampling, raffles, anti-MEV ordering.
sequential work (≈ wall-clock time) elapsed, for fair ordering, timed reveals, proof-of-elapsed-time.
Wedge vs incumbents (e.g. Chainlink VRF): *agent-native* — pay-per-call in USDC, no signup, no KYC, no dashboard, discovered + called over MCP by an autonomous agent, and cheaper (L2 gas). The output is independently verifiable, so the pitch is "verify it yourself," not "trust us."
| capabilityId | price (USD) | input | output | status |
|---|---|---|---|---|
platon.random@v1 | 0.004 | {num_bytes, client_seed?} | {random_hex, proof{state_hash,tick,timestamp,entropy_commitment}, signature} (Ed25519) | live |
platon.beacon@v1 | 0.004 | {} | public round beacon (same proof/sig shape) | live |
platon.commit@v1 | 0.004 | {} | commit half of a commit-reveal | live |
platon.ask@v1 | 0.003 | {question} | entropy-derived answer | live |
chronos.eval@v1 | 0.010 | {seed, difficulty T (1..1e6)} | {scheme, g, y=g^(2^T), proof{pi,l}, modulus} (Wesolowski VDF, RSA-2048) | live |
chronos.verify@v1 | 0.001 | {g, y, T, proof} | {valid} | live |
platon.verify@v1 | ~0.001 | {random_hex, proof, signature} | {valid} | TO BUILD (gap §6) |
Source of truth: oracles/oracles/platon, oracles/oracles/chronos, declared via oracle-core OracleSpec/Capability; invoked today by the relayer (lottery/relayer/ailottery_relayer/oracles.py).
Platon randomness is authenticated by an Ed25519 signature over (random_hex ‖ proof) by the oracle signer; the consumer verifies it against the signer public key published in the signed Hub manifest. The entropy_commitment + state_hash bind the value to the oracle's chaotic state at tick/timestamp (commit-reveal-safe).
Chronos VDF output satisfies y = g^(2^T) mod N with a Wesolowski proof (π, l) checkable in one exponentiation — via chronos.verify@v1 (off-chain) or on-chain (lottery/contracts/src/ChronosVDF.sol, already used by the lottery). The delay is *unforgeable*: producing y requires T sequential squarings (no GPU shortcut), so a valid proof attests that real time/work elapsed.
The product promise: outputs are checkable without trusting the oracle — off-chain in one call, and (Chronos) on-chain in one
verifyEquation.
Two surfaces, same capability:
GET /ai-market/v2/manifest (Ed25519-signed) listsplaton.random@v1, chronos.eval@v1, … with input_schema, output_schema, price_per_call_usd, trust_score; searchable by intent="verifiable randomness".
listed on Glama + MCP registries. This is the agent-native discovery channel — an agent adds the server and immediately sees the tools.
Gaps to close (spec'd in §6): advertise the capabilities + a /ai-market/v2/prices bulk endpoint in a signed /.well-known/ai-market.json; advertise the MCP server there too.
The passthrough MCP server exposes (names chosen for agent legibility; rich param descriptions keep the Glama TDQS score high — see [[glama-mcp-score-maintenance]]):
// tools (each maps 1:1 to an oracle capability via oracle_core.Protocol.invoke)
{ "name": "get_random",
"description": "Unbiasable, Ed25519-signed verifiable randomness. Returns random bytes + a proof you can verify. Pay-per-call (~$0.004 USDC).",
"input_schema": { "num_bytes": "int 1..1024", "client_seed": "optional hex (domain separation)" } }
{ "name": "get_randomness_beacon",
"description": "Latest public randomness beacon (shared across callers in a round). ~$0.004." }
{ "name": "verify_random",
"description": "Verify a Platon randomness output (random_hex+proof+signature) against the published signer key. ~$0.001.",
"input_schema": { "random_hex": "hex", "proof": "object", "signature": "ed25519 sig" } }
{ "name": "compute_vdf",
"description": "Verifiable Delay Function: proves T sequential squarings (≈ elapsed time) over RSA-2048. Returns y=g^(2^T) + Wesolowski proof. ~$0.01.",
"input_schema": { "seed": "hex", "difficulty": "int 1..1000000 (T)" } }
{ "name": "verify_vdf",
"description": "Verify a VDF proof in one exponentiation. ~$0.001.",
"input_schema": { "g": "hex", "y": "hex", "T": "int", "proof": "object" } }Payment is handled by the server under the hood (§5); the calling agent just invokes the tool.
Per-call over the AIMarket v2 micropayment channel (USDC on Base; escrow AIMarketEscrow):
POST /ai-market/v2/channel/open {deposit_usd} → channelId (+24h expiry).POST /ai-market/v2/invoke {capabilityId, input} with X-Payment-Channel: channelId.price_usd.debitChannel(channelId, amount, receiptId, deadline, sig) (EIP-712), the pathwe tested in the journal.
channel/close → settle used to the hub, refund remainder to depositor.Routing fee (routing_fee_bps, default 100 = 1%) is the slice that funds the commons (UBI / oracle pay / gas) — the flywheel fuel. The MCP server opens/settles channels for the agent and must expose its wallet/onboarding (key per chain, USD↔token conversion) — a tracked gap (§6).
Pricing note (competitive): at $0.004/random and $0.01/VDF on cheap L2 settlement, the agent-native offer undercuts human-centric VRF on per-call cost while adding MCP discoverability and self-verify.
platon.verify@v1 — so consumers verify Platon output in one call (currently only Chronoshas verify). [oracle-core capability + handler]
N as hex (not the label "RSA-2048") so VDF verification istrustless without out-of-band knowledge of N.
alexar76/aimarket-oracle-gateway (Glama); FastMCP tools from §4 → Hub invoke./.well-known/ai-market.json that advertises capabilities + a /ai-market/v2/pricesbulk endpoint + the MCP server URL.
get_random, open a channel, callit, receive a signed result, and verify it — paying real USDC, with the routing fee landing in the commons. End-to-end, no human steps.
y verifies on-chain via ChronosVDF.verifyEquation and off-chain via verify_vdf.by itself create buyers. Distribution (Glama, registries, target agent frameworks) is a separate track.