agenthaven-dev-mcp.yml
Agent Bench operates ONE remote MCP server at https://travel.agenthaven.dev/mcp, on the same host as its A2A JSON-RPC endpoint (/a2a) and its documentation. It is POST-only Streamable HTTP without server-initiated streams (a GET returns 405 {"error":{"code":"sse_unsupported","message":"This MCP endpoint does not open server-initiated streams. POST JSON-RPC messages instead."}}), speaks MCP protocol version 2025-06-18, identifies itself as serverInfo {name: agent-bench-travel-merchant, version: 0.1.0}, and answers initialize and tools/list anonymously with two tools carrying full JSON Schema inputSchema AND outputSchema (draft-07), with examples, patterns and const constraints. resources/list and prompts/list return -32601 with a message naming the implemented set: "initialize, ping, tools/list and tools/call". The server is provider-built — the tool names, the merchant_id const "demo-travel-seller" and the mandate/ledger vocabulary match no shared platform fingerprint. The A2A agent card advertises this endpoint in supportedInterfaces[1] and delegates its skill schemas to it. This is the machine-readable contract for the provider: there is no OpenAPI.
One-click install for Cursor, VS Code, Claude, and 20+ other MCP clients, powered by API Commons MCP Install — visit install.apicommons.org for more information.
Endpoint
Hosted endpoint · transport streamable-http
Connect
This is a remote MCP server — point an MCP client at the endpoint URL. Clients with native remote support (Claude, Cursor, VS Code, …):
For stdio-only clients (older Claude Desktop, etc.), bridge with mcp-remote:
If the server requires authentication, add the provider's token/header (e.g. an X-API-Key) per its docs. A quick reachability check:
Tool quality observed · not scored
tools with readOnly / destructive hints
parameters described (14 total)
state-changing tools without destructiveHint
tool-list tokens (2 tools)
tools over ~1,200 tokens
Unhinted state-changing: create_checkout
Heaviest: search_flights 2587, create_checkout 1958
From an anonymous initialize + tools/list capture measured 2026-10-20; tokens are chars/3.6 of each tool definition. Recorded for the Kin Score's agent readiness as an observed signal — it contributes no points yet.
Tools
search_flights— First step. Search flight offers for one route on one date; no authorization needed. Optional airlines keeps only itineraries on those airlines. Returns up to two options, each a quote valid for ten minutes: the cheapest fare and the fastest journey (one option carrying both labels when they coincide), with itinerary — flight numbers, airline, airports, local departure and arrival times, stops — and fare in minor units. Choose one; its id is the input to create_checkout.create_checkout— Second step, after search_flights. quote_id is the id of one of its options (options[].id, q_ followed by a UUID), valid ten minutes; there is no fixed example, and an id search_flights did not mint is refused as quote_unavailable. Needs a bearer authorization (see the agent card's securitySchemes.bearer for where one comes from). Turns the quote into a payment intent and stops there: no ticket is issued. A PaymentIntent is created at Stripe in test mode and confirmed with a test card (pm_card_visa); it succeeds and captures. Test mode: no real card, no real money.
About MCP
The Model Context Protocol (MCP) is an open protocol Anthropic introduced for connecting LLM-based agents to external tools and data sources. Providers publish MCP servers that expose their API surface as structured, discoverable tools — an MCP-compatible client (Claude Desktop, Cursor, Cline, Continue, etc.) can connect to the server and call its tools without any per-provider integration code.
Browse every MCP server on the APIs.io network or compare with the broader Agent Skill surfaces of the same providers.
Work with this as data
Every MCP server here is available over the APIs.io API and to AI agents over MCP.