B2A MCP Server

Adversarial verification + capability discovery for AI outputs.

AI checks its own work and says "looks good" (sycophancy bias). B2A gives you the real score — and fixes it.

The Problem

You: "Is this output good?"
Claude: "Yes, it covers the key points." (~82/100)

You: "Check it with B2A."
B2A:  score: 28/100, grade: FAIL
      issues: [
        "Zero financial metrics, no risk assessment",
        "No evidence for 'expanding' — what expansion?",
        "'Stock might go up' is a tautology, not analysis"
      ]
      gaps: [
        "missing_data: No P/E, revenue, hash rate",
        "missing_analysis: No BTC price scenarios",
        "missing_context: No competitors (RIOT, MARA)"
      ]

Then /enhance fills those gaps with real data and produces a structured analysis. 3 lines in, full report out.

Install (10 seconds)

claude mcp add b2a -- npx b2a-mcp-server

Or with npm:

npm install -g b2a-mcp-server
claude mcp add b2a -- b2a-mcp

Set ANTHROPIC_API_KEY environment variable for verify/enhance tools.

Tools

Tool	Cost	What it does
b2a_discover	Free	Find x402 services for a capability gap
b2a_verify	~$0.01 API	Adversarial audit: score, issues, gaps
b2a_enhance	~$0.05 API	Verify + discover + recommend fixes
b2a_browse	Free	Browse 200+ x402 ecosystem services
b2a_call	Varies	Call any x402 service directly

How it works

Your output (score: 28)
    |
    v
b2a_verify — adversarial audit, sycophancy-free
    |         "score: 28. Here's what's wrong. Here's what's missing."
    v
b2a_enhance — finds services to fill each gap
    |          routes to real-time data, stress-testers, analyzers
    v
Enhanced output (score: 78)
    with real data, structured analysis, verified claims

Why not just ask Claude to check itself?

1. Sycophancy bias — Models trained on human feedback learn to agree. "Is my output good?" → "Yes." Every time.
2. No external data — Claude can't fetch real-time prices, on-chain data, or call paid APIs.
3. No memory — Claude doesn't learn which fixes worked. B2A's feedback loop does.

B2A uses a separate Claude instance with an adversarial "ruthless auditor" mandate. Same model, different role. This breaks the self-assessment loop.

x402 Services (paid, Base USDC)

The MCP server connects to live x402 services:

Service	URL	Price	What it does
Checkpoint402 /verify	checkpoint402.itaeui12.workers.dev	$0.05-0.08	Adversarial verification
Checkpoint402 /enhance	checkpoint402.itaeui12.workers.dev	$0.20-0.70	Verify + enhance with data
Thesis402	thesis402.itaeui12.workers.dev	$0.08	Investment thesis stress-test
data402	data402.itaeui12.workers.dev	Free (100/day)	Real-time crypto, web, on-chain

Ecosystem

200+ services indexed from the x402 ecosystem:

• 75 callable API endpoints
• 68 infrastructure tools
• 30 payment facilitators
• 23 client integrations

Build from source

git clone https://github.com/LTede/b2a-mcp-server.git
cd b2a-mcp-server
npm install
npm run build

License

MIT