How MCP Works
The Model Context Protocol contract that lets Claude, ChatGPT, Gemini, and any MCP client invoke Parlay's typed prediction-market research tools.
The Model Context Protocol (MCP) is the contract that lets an AI client call an external tool server with typed arguments and typed JSON responses. Parlay uses MCP so Claude, ChatGPT, Gemini, OpenClaw, Cursor, Windsurf, and other compatible clients can research prediction markets without a venue-specific SDK.
Current scope
Parlay's MVP is read-only. It helps AI clients search, discover, compare, inspect, scan, and summarize prediction-market data across Polymarket, Kalshi, Limitless, and Manifold.
Why MCP for prediction markets
Connect once
One hosted MCP endpoint exposes the same tool names to every compatible AI client.
Typed tools
Each tool has a JSON schema, so the model sends real arguments and receives predictable fields.
Source-aware results
Responses can include venue names, market URLs, close dates, and caveats that the client can cite.
The MCP descriptor
When a client connects, the server returns a descriptor describing its capabilities. A simplified Parlay descriptor looks like this:
{
"server": {
"name": "Parlay",
"version": "1.0.0",
"transport": "http"
},
"capabilities": {
"tools": {
"search_markets": {
"description": "Search supported prediction-market venues by keyword, topic, venue, or category.",
"inputSchema": {
"type": "object",
"properties": {
"query": { "type": "string" },
"venue": { "type": "string" }
},
"required": ["query"]
}
},
"discover_markets": {
"description": "Browse currently relevant market clusters without starting from an exact keyword.",
"inputSchema": {
"type": "object",
"properties": {
"topic": { "type": "string" },
"venue": { "type": "string" }
}
}
},
"compare_markets": {
"description": "Compare similar questions across supported venues and surface wording differences.",
"inputSchema": {
"type": "object",
"properties": {
"query": { "type": "string" },
"venues": {
"type": "array",
"items": { "type": "string" }
}
},
"required": ["query"]
}
},
"scan_discrepancies": {
"description": "Find notable differences between comparable market questions.",
"inputSchema": {
"type": "object",
"properties": {
"query": { "type": "string" },
"threshold": { "type": "number" }
},
"required": ["query"]
}
},
"inspect_platform": {
"description": "Inspect coverage, status, and known limitations for a supported venue.",
"inputSchema": {
"type": "object",
"properties": {
"venue": { "type": "string" }
},
"required": ["venue"]
}
},
"market_brief": {
"description": "Generate a concise research brief from market results and source links.",
"inputSchema": {
"type": "object",
"properties": {
"query": { "type": "string" },
"include_sources": { "type": "boolean" }
},
"required": ["query"]
}
}
},
"resources": [
{
"uri": "parlay://venues",
"name": "venues",
"description": "Catalog of supported venues with status and tool coverage.",
"mimeType": "application/json"
}
],
"prompts": []
}
}Three primitives carry every MCP server:
Tools
Functions the model can invoke with typed arguments. Parlay exposes six read-only research tools.
Resources
Reference data the client can fetch by URI. Parlay can expose venue coverage metadata this way.
Prompts
Optional reusable prompt templates. Parlay does not depend on prompt templates for the MVP.
How a request flows
Connect to the endpoint
The AI client opens the hosted HTTP MCP endpoint and starts the descriptor handshake.
Server announces capabilities
Parlay returns the tool names, argument schemas, resource URIs, and server metadata the client needs.
Client invokes a tool
The model selects a tool such as search_markets, compare_markets, or market_brief, then sends arguments that match the schema.
Parlay validates account and plan
The MCP Worker checks the signed-in account, plan, quota, and tool access before any venue adapter is called.
Server returns typed JSON
Parlay normalizes venue responses into a consistent JSON shape so the client can summarize the result, cite sources, and ask follow-up questions.
Transport
Parlay's MVP uses hosted HTTP MCP. Client support differs, so use the install guide for your client when it is available.
https://mcp.parlay.run/mcpThe install flow signs you in and links the MCP client to your Parlay account. Your plan and quota are enforced server-side.
Claude is the first installation path Parlay is documenting. ChatGPT connector workflows, Gemini CLI, OpenClaw, Cursor, and Windsurf depend on each client's current MCP support.
No venue credentials required
You do not paste venue secrets into Parlay. The hosted worker handles venue access server-side and returns read-only research data to the MCP client.
Tool example
What a read-only call looks like on the wire:
{
"tool": "market_brief",
"arguments": {
"query": "Fed rate decision markets for June 2026",
"include_sources": true
}
}{
"summary": "Several supported venues list markets related to the June 2026 Fed decision.",
"markets": [
{
"venue": "Polymarket",
"title": "Fed decision market",
"url": "https://example.com/source",
"close_time": "2026-06-17T00:00:00Z"
}
],
"caveats": ["Market wording and settlement criteria can differ by venue."],
"sources": [
{
"venue": "Polymarket",
"url": "https://example.com/source"
}
]
}Not investment advice
Parlay provides research context and source links. It does not provide financial, legal, tax, or investment advice.