x402 Conformance

x402-conformance-mcp

x402 conformance and safety checks as tools, for MCP clients (Claude, Cursor, VS Code, Windsurf and others), the Vercel AI SDK and LangChain. It's a thin entry point into the open x402-conformance suite: the same checks and check IDs as the CLI, the GitHub Action and the test matchers, and the full x402-conformance.report/1 report. It never pays, signs or touches mainnet.

MCP

{ "mcpServers": { "x402-conformance": { "command": "npx", "args": ["-y", "x402-conformance-mcp"] } } }

The same server runs from the Docker image, which starts it on stdio when given no arguments: { "mcpServers": { "x402-conformance": { "command": "docker", "args": ["run", "-i", "--rm", "ghcr.io/atsp-protocol/x402-conformance:1"] } } }. npx -y x402-conformance-mcp --version prints the version, and --help prints the client config and a link to docs/USING.md.

Try asking: "Run the x402 demo", "Check the 402 answer at http://127.0.0.1:4020/paid", or "Start the facilitator double, I'll point my server at it, then probe http://127.0.0.1:4020/paid".

Vercel AI SDK

import { generateText } from "ai";

import { x402ConformanceTools } from "x402-conformance-mcp/ai-sdk";

await generateText({ model, tools: x402ConformanceTools(), prompt: "Is the 402 answer at https://api.example.com/paid valid x402?" });

LangChain and LangGraph

import { createReactAgent } from "@langchain/langgraph/prebuilt";

import { x402ConformanceTools } from "x402-conformance-mcp/langchain";

const agent = createReactAgent({ llm, tools: x402ConformanceTools() });

The AI SDK and LangChain adapters give an agent the four read-only tools (check, demo, explain, share). The double and probe tools stay in MCP, where a person is driving. The adapters need ai 5.0.0 or later, or @langchain/core 0.3.0 or later, installed next to them; both are optional peer dependencies.

The corpus

A failing report can become an anonymised case in the public corpus of real x402 failures. x402_share_case prepares it and returns the link; the user reviews it and submits it, or doesn't. Every report names the entry point that produced it (via: mcp, ai-sdk or langchain) in the report and the User-Agent; there's no other identifier and no telemetry.

Apache-2.0. Ryan Gillespie / Optitransfer.