preflight402
MCP server that preflights endpoints before agent payments, returning a trust verdict with proceed/caution/avoid recommendation.
README
preflight402
One free call before your agent pays. Health, authenticity, and Sybil-filtered reputation — one verdict.
A free trust/health preflight for the agent payment economy (x402). It probes
an endpoint before your agent pays and returns one trust-preview.v1 verdict:
liveness, TLS, the 402 handshake (x402 v1/v2 + MPP detection), price sanity,
continuous uptime history, ERC-8004 identity binding, and Sybil-filtered
on-chain reputation — rolled into a proceed / caution / avoid recommendation
with plain-language reasons. No wallet, no key, no charge.
Why filtered reputation matters: one live agent shows a 99.8/100 average from 996 reviewers — until Sybil filtering collapses those reviewers into 12 independent funding clusters at 89.7. Raw reputation is trivially farmed; this is what the verdict actually scores. (Separately, only ~4% of x402 payees bind to any ERC-8004 identity at all — so most verdicts run on health, price, and handshake, and say so honestly.)
Why "listed" means little: across 4.98M probes over 9.24 days of a 51,331-endpoint catalog, two squatting hosts account for 58% of every listed x402 endpoint — one serving a uniform 404 behind a "This app isn't live yet" placeholder, the other a uniform Cloudflare TLS failure. Strip them out and 78% of real endpoints work; measured per provider, about one host in seven serves no valid 402. Full methodology, corrections and caveats: docs/checkpoint-m3.md — including why a GET-only crawl (ours included, before correction) overstates x402 invalidity ~2×.
30-second quickstart
Point any agent at the hosted MCP endpoint — no wallet, no key, no install:
https://preflight402.ironshell.io/mcp
Or check an endpoint over plain HTTP:
curl 'https://preflight402.ironshell.io/preflight?url=https://api.example.com/paid'
Guard every payment automatically
preflight402-guard turns the service into a payment gate for the
x402 Python SDK — safety becomes
default-on instead of something an agent has to remember to call:
pip install "preflight402-guard[x402]" # PyPI publish pending; for now: pip install "git+https://github.com/duskwire/preflight402.git#subdirectory=guard"
from preflight402_guard import Guard
from x402 import x402Client
guard = Guard() # block "avoid", warn on "caution"
client = x402Client()
guard.install(client) # every payment is preflighted before signing;
# a bad verdict raises PaymentAbortedError
It also cross-checks that the payee your client selected matches the endpoint
that was preflighted (the 402's resource URL is attacker-controlled), enforces
an optional max_price_usd ceiling against the actual selected terms, and
fails open by default so your commerce never depends on our uptime. There's
a CLI too: preflight402-guard check <url>. See guard/README.md.
Status
Live at preflight402.ironshell.io and in
the official MCP registry as
io.ironshell/preflight402. The free preflight engine (health + 402 parse +
verdict), continuous probing with uptime history, ERC-8004 binding, and the
Sybil filter are all shipped and serving live. The whole service is free —
there is no paywall.
Use it
As an MCP tool (no wallet, no key)
The preflight tool takes a url and returns a trust-preview.v1 verdict.
Easiest — point any MCP client at the hosted instance, no install:
https://preflight402.ironshell.io/mcp (streamable-http)
Or run it yourself over stdio. Claude Code — one line (PyPI publish pending;
until then point --directory at a clone):
claude mcp add preflight402 -- uvx --from preflight402 preflight402-mcp
# pre-PyPI: claude mcp add preflight402 -- uv run --directory /path/to/preflight402 preflight402-mcp
Claude Desktop — add to claude_desktop_config.json:
{
"mcpServers": {
"preflight402": {
"command": "uvx",
"args": ["--from", "preflight402", "preflight402-mcp"]
}
}
}
Either way, run it as a hosted HTTP server with
preflight402-mcp --transport streamable-http (serves the same tool at
http://<host>:8000/mcp).
As a REST call
The hosted instance also serves REST:
curl 'https://preflight402.ironshell.io/preflight?url=https://api.example.com/paid'
Or run it yourself (serves REST + MCP on one port):
uv run uvicorn preflight402.api.app:app --port 8402
curl 'http://localhost:8402/preflight?url=https://api.example.com/paid'
Development
Requires uv.
uv sync # create venv + install deps
uv run uvicorn preflight402.api.rest:app # serve on :8000
curl http://localhost:8000/healthz # {"status":"ok","version":"0.1.0"}
uv run pytest # tests
uv run ruff check . # lint
uv run ruff format --check . # formatting
Layout
src/preflight402/
├── api/ # REST + MCP server
├── probe/ # async prober, TLS inspection, 402 parsers (x402 v1/v2, MPP)
├── verdict/ # rules -> trust-preview.v1 JSON
├── chains/ # ChainVerifier interface: EVM (Base), SVM (Solana)
├── reputation/ # ERC-8004 subgraph client, endpoint binding, Sybil filter
├── ingest/ # endpoint seed ingesters (Bazaar, x402scan, ...)
├── scheduler/ # probe loop with per-host politeness
└── db/ # SQLite (WAL) schema + queries
guard/ # preflight402-guard: client-side auto-preflight for the x402 SDK
tests/ # unit/ + golden/ (captured 402 responses) + integration/ (marked slow)
deploy/ # Dockerfile + deploy notes
docs/ # trust-preview.v1 schema + API docs (M8)
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。