Purpose
aimarket-oracle-gateway is an MCP (Model Context Protocol) server that serves as the discovery and consumption storefront for the oracle-as-a-service specification. It is aimed at AI agents, provides verifiable oracle services, and implements pay-per-call via the AIMarket protocol. Every result can be independently verified, with verification methods including signatures, VDF proofs, classical certificates, and deterministic replay.
Main capabilities
The server provides seventeen oracle families, with 35 tools in total. Some of the oracle services and verification methods explicitly listed in the material include:
- Platon: Ed25519 signature over (random_hex ‖ proof).
- Chronos: Wesolowski VDF proof, supporting low-cost verify_vdf.
- LUMEN: re-derive PageRank from a committed graph.
- Murmuration: deterministic re-aggregation based on input values.
- Landauer: bit-by-bit erasure counts from the operation DAG and circuit_commitment.
- Fermat: check the dual certificate T(v) in O(E).
- Ablation: sandpile replay, topple_total + tau is order-independent.
- Lattice / Turing: determined by (count, dim, skip) / seed respectively.
- Colony: acceptable lower bound lower_bound plus gap certificate.
- Percola: recomputable percolation sweep plus f_c.
- Sortes: ECVRF 80-byte proof, verifiable offline by anyone (RFC 9381).
- Gauss: deterministic replay of GP posterior and Expected-Improvement suggestions.
- Aestus: RSW time-lock puzzle can be re-derived; the modulus factorization is destroyed at each sealing.
- Betti: recomputable Vietoris-Rips persistence diagram plus bottleneck distance.
- Kantor: check Kantorovich dual potentials in O(m*n), without re-solving.
- Fourier: recomputable Laplacian spectrum plus Fiedler value lambda-2.
Usage
The transport is stdio (aimarket_oracle_gateway/stdio_server.py). Built on the official Model Context Protocol Python SDK (mcp, FastMCP). Compatible hosts include Claude Desktop, Cursor, Glama, and any MCP client that supports a stdio server.
Configuration is done via environment variables. If neither URL is set, the server fails closed with an explicit message.
For payment security, according to specification 03, spending limits are enforced on the client side and cannot be overridden by a prompt-injected agent.
Claude Desktop is configured by adding the corresponding configuration to the mcpServers entry.
The first call is usually list_oracle_capabilities; high-traffic tools also announce monthly bundles in that tool. Pay-per-call remains the default, and bundles are Hub-side packaging.
Release channel: listed on Glama as glama.ai/mcp/servers/alexar76/aimarket-oracle-gateway.
Applicable scenarios
Suitable for AI agent workflows that require verifiable computation results, such as scenarios requiring verifiable randomness, verifiable delay functions, decentralized reputation scoring, deterministic replay verification, and so on. Suitable for integration with agent hosts that support the MCP protocol, calling oracle services on demand via the AIMarket protocol.
Limitations
This repository is a read-only mirror, published from the specification AI-Factory monorepo. Pull requests are not accepted, and any commits pushed here will be overwritten by scripts/mirror_satellites.sh on the next sync. To report bugs or make requests, please open an issue.
Regarding specific subscription prices, bundle fees, detailed parameters of each oracle, supported blockchain networks, environment variable names, and complete value examples, the material does not currently provide them.
