A community trust scoring layer that runs on top of any MCP server in the ecosystem. It watches listed servers for behavioral drift, dependency changes, or permission creep, then assigns an A through F grade you can check before connecting. The oracle itself is independent and explicitly not for sale, so the incentive stays aligned with accuracy rather than promotion. If you're wiring up third-party MCP servers in production or you maintain a client that auto-discovers servers, this gives you a programmatic trust signal without auditing every repo yourself. Free to query, runs over streamable HTTP, and the grades update as the underlying servers change.
Add any website to Claude, Cursor, or Codex as agent-callable MCP tools — in one line of config. No per-server build, no SDK.
▶ Watch the 30-sec demo — paste a URL → tools appear → agent calls add_to_cart.
wmcp.sh is a real, spec-compliant MCP server (JSON-RPC 2.0 over Streamable HTTP) that turns any URL into agent-callable tools — Shopify product pages, OpenAPI specs, JSON-LD pages, OAuth-protected APIs. Every other MCP server is one-per-service and hand-built; wmcp is one endpoint that adapts to whatever URL you point it at. Paste a config line and your agent can read prices, list variants, or call an API on a site that ships no MCP server of its own.
Live now: wmcp.sh · Directory · Blog · /llms.txt
Point any MCP client — Claude Code, Claude Desktop, Cursor, Codex, Cline, VS Code — at a per-URL endpoint: /mcp/u/<base64url-of-your-url>. That's the whole install. The server exposes that page's extracted tools natively (tools/list + tools/call):
{
"mcpServers": {
"allbirds": {
"type": "http",
"url": "https://wmcp.sh/mcp/u/aHR0cHM6Ly93d3cuYWxsYmlyZHMuY29tL3Byb2R1Y3RzL21lbnMtd29vbC1ydW5uZXJz"
}
}
}
That base64url string decodes to https://www.allbirds.com/products/mens-wool-runners. Encode your own:
node -e "process.stdout.write(Buffer.from('https://your-url.example').toString('base64url'))"
Compose several sites into one MCP server — repeat the url param; tool names are namespaced per host (site_a__add_to_cart):
{ "mcpServers": { "wmcp": { "type": "http", "url": "https://wmcp.sh/mcp/url?url=https://site-a.example&url=https://site-b.example" } } }
tools/list (discovery) is free. tools/call (live execution) needs a paid key, passed as a Bearer header:
{ "mcpServers": { "wmcp": { "type": "http", "url": "https://wmcp.sh/mcp/u/<base64url>", "headers": { "Authorization": "Bearer webmcp_live_…" } } } }
type may be "http" or "streamable-http" (same transport). For Claude Desktop builds without remote-MCP support, bridge with npx mcp-remote <url>. Full walkthrough: /how-to/install-claude-desktop-mcp.
# any URL → MCP-shaped tool list
curl 'https://wmcp.sh/api/v1/tools?url=https://www.allbirds.com/products/mens-wool-runners'
Returns a JSON tool list in the Claude tool_use / OpenAI function_call / MCP shape — get_price, add_to_cart, list_variants — that any agent can invoke.
A 5-tier extraction chain turns most URLs into tools with zero per-site work. Each tier is tried in order; the first that detects the URL wins. The last tier is an LLM fallback that produces best-effort tools and caches the result.
┌──────────────────────────────────────────────────┐
│ URL in → │
│ 1. Shopify (~4M storefronts, <50ms) │
│ 2. JSON-LD (schema.org Product, etc.) │
│ 3. OpenAPI (any 3.x / Swagger 2.0 → tools)│
│ 4. Provider (Stripe, GitHub, Slack…) │
│ 5. LLM-fallback (Claude Haiku, cached 30d) │
│ → MCP tools out │
└──────────────────────────────────────────────────┘
Cached on Cloudflare KV; sub-50ms on a cache hit. Discovery (tools/list) is free; live execution (tools/call) runs through the same engine and needs a paid key. For sites that block bot fetches or hydrate client-side from authed APIs, a Chrome extension extracts schemas in your browser and pushes them back to the shared cache. Full design in docs/ARCHITECTURE.md.
Today, "an MCP server for X" means someone wrote, built, and deployed a dedicated server for service X. There are hundreds of them, each one a separate repo and a separate install.
wmcp inverts that:
| Hand-built MCP server | wmcp.sh | |
|---|---|---|
| Coverage | One service per server | Any URL, one endpoint |
| Setup | Clone, build, configure, deploy | Paste one config line |
| New target | Write & ship a new server | Encode the URL — done |
| Auth'd upstreams | Bake credentials into each server | Shared OAuth/credential vault proxy at /mcp/<provider> |
| Discovery | Find/trust each repo | Public directory of what's already cached |
There's also a saved toolset endpoint — /mcp/set/<id> — that bundles a curated set of URLs behind one stable MCP server (a Pro feature). It's genuinely novel: instead of one-per-service, wmcp turns arbitrary URLs into MCP tools on demand.
https://wmcp.sh/mcp/<provider> (Google, GitHub, Slack, Notion, Linear, Discord, Stripe, Anthropic).pip install wmcp. Native fetch-style client, zero required deps.npm install @wmcp/sdk with one-line adapters for @wmcp/sdk/anthropic, @wmcp/sdk/openai, @wmcp/sdk/langchain, and @wmcp/sdk/vercel-ai.import { WmcpClient } from "@wmcp/sdk";
import { toAnthropicTools } from "@wmcp/sdk/anthropic";
const client = new WmcpClient(); // anonymous = 100 reads/day per IP
const tools = toAnthropicTools(await client.tools("https://www.allbirds.com/products/mens-wool-runners"));
Two separate ladders. Self-serve API plans for anyone, and a done-for-you service if you want a custom adapter built and run for you.
Self-serve API (manage at /dashboard):
| Tier | Cost | What you get |
|---|---|---|
| Free | $0 | tools/list + REST reads, IP-rate-limited, no live execute |
| Builder | $39/mo | Entry paid tier: live tools/call execute, OAuth vault, higher quotas |
| Pro | $99/mo | High quotas, priority cache, full execute + vault |
| Reseller | $299/mo | High-volume quotas, email support |
Done-for-you service (/managed):
| Tier | Cost | What you get |
|---|---|---|
| Starter | $499 one-time | Custom adapter built for your site + verified directory badge |
| Managed Retainer | $999/mo | Ongoing maintenance, featured listing, white-label MCP at mcp.yourbrand.com |
| Enterprise | $4,999+/mo | Self-hosted deployment + SLA + dedicated support |
.
├── adapters/ Shared adapter library (worker + extension import the same files)
│ ├── shopify.js ~4M storefronts, read + add_to_cart
│ ├── jsonld.js generic schema.org Product
│ ├── openapi.js any OpenAPI 3.x / Swagger 2.0 → tools
│ ├── llm.js Claude Haiku fallback
│ ├── CONTRACT.md the detect() / extract() / actions interface
│ ├── _template/ starter scaffold for new adapters
│ └── _test/run.mjs node:test harness, zero deps
│
├── worker/ Hosted API + MCP server at wmcp.sh
│ └── src/ Hono routes, mcp_server.ts, OAuth, billing, SEO pages, token vault
│
├── extension/ Chrome MV3 extension (cache-back path)
│
├── sdks/
│ ├── python/wmcp/ pip install wmcp
│ └── javascript/src/ @wmcp/sdk (+ anthropic / openai / langchain / vercel-ai adapters)
│
├── docs/ QUICKSTART · ARCHITECTURE · API · SELF_HOSTING
├── launch/ Launch artifacts (HN draft, registry submissions, demo)
├── AGENTS.md rules for AI coding agents working on this repo
├── CONTRIBUTING.md how to send a PR
├── ADAPTERS_WANTED.md open targets, ranked by difficulty
└── LICENSE MIT
New adapters are welcome via PR. The whole point is that adding a site shouldn't mean building a new server.
git clone https://github.com/New1Direction/webmcp-anything
cd webmcp-anything
cp -r adapters/_template adapters/mysite
# edit adapters/mysite/adapter.js — implement detect() + extract()
node adapters/_test/run.mjs
Adapter files are plain ES modules with zero npm deps so the same file runs in Cloudflare Workers, Chrome MV3 service workers, and Node 18+. Read CONTRIBUTING.md, the adapter contract, and open targets in ADAPTERS_WANTED.md. If you're an AI coding agent, read AGENTS.md first.
/api/v1/tools)pip install wmcp) + JS (@wmcp/sdk) SDKs and a Chrome extensionlaunch/registry/SUBMISSION_PLAYBOOK.mdFound a bug? Open an issue. Have a site we should support? ADAPTERS_WANTED.md.
MIT. See LICENSE. Repo: github.com/New1Direction/webmcp-anything.